Engineering Optimization, 2nd Ed, Wiley (2006)

ENGINEERING OPTIMIZATION

Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

ENGINEERING OPTIMIZATION Methods and Applications SECOND EDITION

A. Ravindran Industrial and Manufacturing Engineering Pennsylvania State University

K. M. Ragsdell Engineering Management and Systems Engineering University of Missouri–Rolla

G. V. Reklaitis Chemical Engineering Purdue University

John Wiley & Sons, Inc.

This book is printed on acid-free paper.嘷 ⬁ Copyright 䉷 2006 by John Wiley & Sons. All rights reserved Published by John Wiley & Sons, Inc., Hoboken, New Jersey Published simultaneously in Canada No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 646-8600, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at http: / / www.wiley.com / go / permission. Limit of Liability / Disclaimer of Warranty: While the publisher and the author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor the author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. For general information about our other products and services, please contact our Customer Care Department within the United States at (800) 762-2974, outside the United States at (317) 572-3993 or fax (317) 572-4002. Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic books. For more information about Wiley products, visit our web site at www.wiley.com. Library of Congress Cataloging-in-Publication Data: Ravindran, A., 1944– Engineering optimization: methods and applications / A. Ravindran, K. M. Ragsdell, G. V. Reklaitis.—2nd ed. p. cm. Includes bibliographical references and indexes. ISBN-13 978-0-471-55814-9 (cloth) ISBN-10 0-471-55814-1 (cloth) 1. Engineering—Mathematical models. 2. Mathematical optimization. I. Ragsdell, K. M. II. Reklaitis, G. V., 1942– III. Title. TA342.R44 2006 620⬘.0042—dc22 2005044611 Printed in the United States of America 10

9

8

7

6

5

4

3

2

1

CONTENTS

Preface 1

xiii

Introduction to Optimization

1.1

Requirements for the Application of Optimization Methods / 2 1.1.1 Defining the System Boundaries / 2 1.1.2 Performance Criterion / 3 1.1.3 Independent Variables / 4 1.1.4 System Model / 5

1.2

Applications of Optimization in Engineering / 6 1.2.1 Design Applications / 8 1.2.2 Operations and Planning Applications / 15 1.2.3 Analysis and Data Reduction Applications / 20 1.2.4 Classical Mechanics Applications / 26 1.2.5 Taguchi System of Quality Engineering / 27

1.3

Structure of Optimization Problems / 28

1.4

Scope of This Book / 29

1

References / 30 2

Functions of a Single Variable

2.1

Properties of Single-Variable Functions / 32

2.2

Optimality Criteria / 35

2.3

Region 2.3.1 2.3.2 2.3.3

2.4

Polynomial Approximation or Point Estimation Methods / 55 2.4.1 Quadratic Estimation Methods / 56

32

Elimination Methods / 45 Bounding Phase / 46 Interval Refinement Phase / 48 Comparison of Region Elimination Methods / 53

v

vi

CONTENTS

2.4.2

Successive Quadratic Estimation Method / 58

2.5

Methods Requiring Derivatives / 61 2.5.1 Newton–Raphson Method / 61 2.5.2 Bisection Method / 63 2.5.3 Secant Method / 64 2.5.4 Cubic Search Method / 65

2.6

Comparison of Methods / 69

2.7

Summary / 70 References / 71 Problems / 71

3

Functions of Several Variables

3.1

Optimality Criteria / 80

3.2

Direct-Search Methods / 84 3.2.1 The S 2 (Simplex Search) Method / 86 3.2.2 Hooke–Jeeves Pattern Search Method / 92 3.2.3 Powell’s Conjugate Direction Method / 97

3.3

Gradient-Based Methods / 108 3.3.1 Cauchy’s Method / 109 3.3.2 Newton’s Method / 111 3.3.3 Modified Newton’s Method / 115 3.3.4 Marquardt’s Method / 116 3.3.5 Conjugate Gradient Methods / 117 3.3.6 Quasi-Newton Methods / 123 3.3.7 Trust Regions / 127 3.3.8 Gradient-Based Algorithm / 128 3.3.9 Numerical Gradient Approximations / 129

3.4

Comparison of Methods and Numerical Results / 130

3.5

Summary / 137

78

References / 137 Problems / 141 4

Linear Programming

4.1

Formulation of Linear Programming Models / 149

149

CONTENTS

vii

4.2

Graphical Solution of Linear Programs in Two Variables / 154

4.3

Linear Program in Standard Form / 158 4.3.1 Handling Inequalities / 159 4.3.2 Handling Unrestricted Variables / 159

4.4

Principles of the Simplex Method / 161 4.4.1 Minimization Problems / 172 4.4.2 Unbounded Optimum / 173 4.4.3 Degeneracy and Cycling / 174 4.4.4 Use of Artificial Variables / 174 4.4.5 Two-Phase Simplex Method / 176

4.5

Computer Solution of Linear Programs / 177 4.5.1 Computer Codes / 177 4.5.2 Computational Efficiency of the Simplex Method / 179

4.6

Sensitivity Analysis in Linear Programming / 180

4.7

Applications / 183

4.8

Additional Topics in Linear Programming / 183 4.8.1 Duality Theory / 184 4.8.2 Dual Simplex Method / 188 4.8.3 Interior Point Methods / 189 4.8.4 Integer Programming / 205 4.8.5 Goal Programming / 205

4.9


5

Constrained Optimality Criteria

5.1

Equality-Constrained Problems / 218

5.2

Lagrange Multipliers / 219

5.3

Economic Interpretation of Lagrange Multipliers / 224

5.4

Kuhn–Tucker Conditions / 225 5.4.1 Kuhn–Tucker Conditions or Kuhn–Tucker Problem / 226 5.4.2 Interpretation of Kuhn–Tucker Conditions / 228

5.5

Kuhn–Tucker Theorems / 229

218

viii

CONTENTS

5.6

Saddlepoint Conditions / 235

5.7

Second-Order Optimality Conditions / 238

5.8

Generalized Lagrange Multiplier Method / 245

5.9

Generalization of Convex Functions / 249

5.10


6

Transformation Methods

6.1

Penalty Concept / 261 6.1.1 Various Penalty Terms / 262 6.1.2 Choice of Penalty Parameter R / 277

6.2

Algorithms, Codes, and Other Contributions / 279

6.3

Method 6.3.1 6.3.2 6.3.3 6.3.4 6.3.5 6.3.6 6.3.7 6.3.8

6.4

Summary / 293

260

of Multipliers / 282 Penalty Function / 283 Multiplier Update Rule / 283 Penalty Function Topology / 284 Termination of the Method / 285 MOM Characteristics / 286 Choice of R-Problem Scale / 289 Variable Bounds / 289 Other MOM-Type Codes / 293


Constrained Direct Search

7.1

Problem Preparation / 306 7.1.1 Treatment of Equality Constraints / 306 7.1.2 Generation of Feasible Starting Points / 309

7.2

Adaptations of Unconstrained Search Methods / 309 7.2.1 Difficulties in Accommodating Constraints / 310 7.2.2 Complex Method / 312 7.2.3 Discussion / 320

305

ix

CONTENTS

7.3

Random-Search Methods / 322 7.3.1 Direct Sampling Procedures / 322 7.3.2 Combined Heuristic Procedures / 326 7.3.3 Discussion / 329

7.4


8

Linearization Methods for Constrained Problems

8.1

Direct 8.1.1 8.1.2 8.1.3

8.2

Separable Programming / 359 8.2.1 Single-Variable Functions / 359 8.2.2 Multivariable Separable Functions / 362 8.2.3 Linear Programming Solutions of Separable Problems / 364 8.2.4 Discussion and Applications / 368

8.3

Summary / 372

336

Use of Successive Linear Programs / 337 Linearly Constrained Case / 337 General Nonlinear Programming Case / 346 Discussion and Applications / 355


Direction Generation Methods Based on Linearization

378

9.1

Method 9.1.1 9.1.2 9.1.3

of Feasible Directions / 378 Basic Algorithm / 380 Active Constraint Sets and Jamming / 383 Discussion / 387

9.2

Simplex Extensions for Linearly Constrained Problems / 388 9.2.1 Convex Simplex Method / 389 9.2.2 Reduced Gradient Method / 399 9.2.3 Convergence Acceleration / 403

9.3

Generalized Reduced Gradient Method / 406 9.3.1 Implicit Variable Elimination / 406 9.3.2 Basic GRG Algorithm / 410

x

CONTENTS

9.3.3 9.3.4

Extensions of Basic Method / 419 Computational Considerations / 427

9.4

Design 9.4.1 9.4.2 9.4.3

Application / 432 Problem Statement / 433 General Formulation / 434 Model Reduction and Solution / 437

9.5


10

Quadratic Approximation Methods for Constrained Problems 450

10.1

Direct Quadratic Approximation / 451

10.2

Quadratic Approximation of the Lagrangian Function / 456

10.3

Variable Metric Methods for Constrained Optimization / 464

10.4

Discussion / 470 10.4.1 Problem Scaling / 470 10.4.2 Constraint Inconsistency / 470 10.4.3 Modification of H (t) / 471 10.4.4 Comparison of GRG with CVM / 471

10.5


11

Structured Problems and Algorithms

11.1

Integer 11.1.1 11.1.2 11.1.3

11.2

Quadratic Programming / 494 11.2.1 Applications of Quadratic Programming / 494 11.2.2 Kuhn–Tucker Conditions / 498

11.3

Complementary Pivot Problems / 499

11.4

Goal Programming / 507

11.5

Summary / 518

481

Programming / 481 Formulation of Integer Programming Models / 482 Solution of Integer Programming Problems / 484 Guidelines on Problem Formulation and Solution / 492

CONTENTS

xi


Comparison of Constrained Optimization Methods

12.1

Software Availability / 530

12.2

A Comparison Philosophy / 531

12.3

Brief History of Classical Comparative Experiments / 533 12.3.1 Preliminary and Final Results / 535

12.4

Summary / 539

530

References / 539 13

Strategies for Optimization Studies

13.1

Model Formulation / 543 13.1.1 Levels of Modeling / 544 13.1.2 Types of Models / 548

13.2

Problem Implementation / 552 13.2.1 Model Assembly / 553 13.2.2 Preparation for Solution / 554 13.2.3 Execution Strategies / 580

13.3

Solution Evaluation / 588 13.3.1 Solution Validation / 589 13.3.2 Sensitivity Analysis / 590

13.4

Summary / 594

542


Engineering Case Studies

14.1

Optimal Location of Coal-Blending Plants by Mixed-Integer Programming / 603 14.1.1 Problem Description / 604 14.1.2 Model Formulation / 604 14.1.3 Results / 609

14.2

Optimization of an Ethylene Glycol–Ethylene Oxide Process / 610 14.2.1 Problem Description / 610

603

xii

CONTENTS

14.2.2 14.2.3 14.2.4

Model Formulation / 612 Problem Preparation / 618 Discussion of Optimization Runs / 618

14.3

Optimal Design of a Compressed Air Energy Storage System / 621 14.3.1 Problem Description / 621 14.3.2 Model Formulation / 622 14.3.3 Numerical Results / 627 14.3.4 Discussion / 629

14.4

Summary / 630 References / 631

Appendix A

Review of Linear Algebra

A.1

Set Theory / 633

A.2

Vectors / 633

A.3

Matrices / 634 A.3.1 Matrix Operations / 635 A.3.2 Determinant of a Square Matrix / 637 A.3.3 Inverse of a Matrix / 637 A.3.4 Condition of a Matrix / 639 A.3.5 Sparse Matrix / 639

A.4

Quadratic Forms / 640 A.4.1 Principal Minor / 641 A.4.2 Completing the Square / 642

A.5

Convex Sets / 646

633

Appendix B

Convex and Concave Functions

648

Appendix C

Gauss–Jordan Elimination Scheme

651

Author Index

653

Subject Index

659

PREFACE We are deeply gratified by the enthusiastic response given to the first edition of our book by our colleagues and students. We took great care in preparing the second edition. We added some new material, revised the explanation and presentation of some topics, and increased the number of problems. Examples of new material include more application of optimization problems in Chapter 1, discussion of duality theory and interior point methods for solving LP problems in Chapter 4, new sections on the generalized Lagrange multiplier method and generalization of convex functions in Chapter 5, a new section on goal programming for solving multiobjective optimization problems in Chapter 11, and the inclusion of software availability for solving nonlinear programs in Chapter 12. To make room for the new material and to keep the size of the book down, we deleted discussion of cutting plane methods in Chapter 8, gradient projection methods, in Chapter 9 and geometric programming in Chapter 11. This is a text on the practical aspects of optimization methodology, with a major focus on the techniques and stratagems relevant to engineering application arising in design, operations, and analysis. Attention is given primarily to techniques applicable to problems in continuous variables that involve realvalued constraint functions and a single real-valued objective function. In short, we treat the methodology often categorized as nonlinear programming. Within this framework a broad survey is given of all-important families of optimization methods, ranging from those applicable to the minimization of a single-variable function to those most suitable for large-scale nonlinear constrained problems. Included are discussions not only of classical methods, important for historical reasons and for their role in motivating subsequent developments in this field, but also of promising new techniques, such as those of successive quadratic programming and goal programming. Our development is aimed at imparting an understanding of the logic of the methods, of the key assumptions that underlie them, and of the comparative merits of the methods. Proofs and mathematical derivations are given only if they serve to explain key steps or properties of algorithms. Generally, we simply cite the original literature source for proofs and use the pages of this book to motivate and explain the key steps underlying the mathematical constructions. Thus, our aim is to inform the engineer-user of optimization methodology rather than to prepare the software specialist who will develop computer implementations of the algorithms. In keeping with this aim, we xiii

xiv

PREFACE

have given considerable attention to practical issues such as model formulation, implementation, preparation for solution, starting point generation, and the selection of execution strategies. A major chapter (Chapter 13) is devoted to strategies for carrying out optimization studies; another (Chapter 12) reviews the state-of-the-art optimization software and the results of existing comparative studies for solving nonlinear programs; and a third (Chapter 14) discusses three significant engineering case studies. In addition, a considerable fraction of each chapter is allocated to engineering examples drawn from the chemical, industrial, and mechanical engineering backgrounds of the authors. While a number of excellent books are available that deal in detail with the rich theoretical and numerical analysis issues that are relevant to nonlinear programming, this book is unique in the features outlined above: broad treatment of up-to-date method; conceptual, rather than formal, presentation; and focus on issues relevant to engineering studies. The first edition of the book was developed over a period of eight years in which various drafts were used in a one-semester interdisciplinary engineering optimization course team-taught by the authors to senior undergraduate and first-year graduate students at Purdue University. For these students, this course was typically the first systematic exposure to optimization methods. The students’ mathematical preparation consisted of the calculus and linear algebra coursework typical of BS engineering curricula; hence, that is all that is required of the reader of this book. The organization of the book has also benefited considerably from the authors’ experience in teaching a television course on engineering optimization broadcast to regional campuses, nonresident MS, and professional engineering audiences. The authors have used the first edition as a text in engineering optimization courses taught at Arizona, Oklahoma, and Purdue universities. The book has also been used in in-plant training courses, with National Technological University courses, and courses taught live over the Internet. We are therefore confident that the book can serve as a text for conventional classroom lectures, for television courses, for on-line courses, and for industrial short courses as well as for self-study. Two different course organizations have been followed in teaching from this text: an all-lecture, one-semester course involving 45 fifty-minute lectures and a lecture–recitation format involving 30 lectures and 15 recitation– discussion sessions. In the former case, the entire contents of the book, except for Chapter 14, can be covered in numerical chapter sequence. In the latter case, Chapters 1, 13, and 14, as well as additional case studies and examples, are discussed in the recitation sessions, while the methodology chapters (Chapters 2–10) and Chapter 12 are covered in the lectures. In this format, Chapter 11 was omitted because of the limited lecture hours. Homework problems in both formats should include problems and computer exercises given at the end of the chapters. Student computer solutions can be carried out using the software programs referenced in Chapter 12. The development and evolution of this book have benefited substantially from the advice and counsel, both conscious and unintentional, of our col-

PREFACE

xv

leagues and mentors in our respective engineering disciplines, who are too numerous to acknowledge individually. We are indebted for the numerous questions and pleas for clarification raised by our students in engineering optimization, who have persevered through various revisions of the manuscript. Their persistent, often biting, and usually justifiable criticism, overt and sub rosa, has been a key driving force in the revision process. We wish to thank our Wiley editor, Robert Argentieri, for his perseverance and patience. We express our sincere appreciation to Ajay Natarajan, an industrial engineering doctoral student at Penn State University, for his careful preparation of the author and subject indexes for the second edition. Finally we are grateful to the instructors who have adopted our first edition and for their encouragement and helpful suggestions that made the second edition a reality. University Park, Pennsylvania Rolla, Missouri West Lafayette, Indiana

A. (RAVI) RAVINDRAN K. M. RAGSDELL G. V. REKLAITIS

1 INTRODUCTION TO OPTIMIZATION This text is an introduction to optimization theory and its application to problems arising in engineering. In the most general terms, optimization theory is a body of mathematical results and numerical methods for finding and identifying the best candidate from a collection of alternatives without having to explicitly enumerate and evaluate all possible alternatives. The process of optimization lies at the root of engineering, since the classical function of the engineer is to design new, better, more efficient, and less expensive systems as well as to devise plans and procedures for the improved operation of existing systems. The power of optimization methods to determine the best case without actually testing all possible cases comes through the use of a modest level of mathematics and at the cost of performing iterative numerical calculations using clearly defined logical procedures or algorithms implemented on computing machines. The development of optimization methodology will therefore require some facility with basic vector–matrix manipulations, a bit of linear algebra and calculus, and some elements of real analysis. We use mathematical concepts and constructions not simply to add rigor to the proceedings but because they are the language in terms of which calculation procedures are best developed, defined, and understood. Because of the scope of most engineering applications and the tedium of the numerical calculations involved in optimization algorithms, the techniques of optimization are intended primarily for computer implementation. However, although the methodology is developed with computers in mind, we do not delve into the details of program design and coding. Instead, our emphasis is on the ideas and logic underlying the methods, on the factors involved in selecting the appropriate techniques, and on the considerations important to successful engineering application. Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

1

2

INTRODUCTION TO OPTIMIZATION

1.1 REQUIREMENTS FOR THE APPLICATION OF OPTIMIZATION METHODS

To apply the mathematical results and numerical techniques of optimization theory to concrete engineering problems, it is necessary to clearly delineate the boundaries of the engineering system to be optimized, to define the quantitative criterion on the basis of which candidates will be ranked to determine the ‘‘best,’’ to select the system variables that will be used to characterize or identify candidates, and to define a model that will express the manner in which the variables are related. This composite activity constitutes the process of formulating the engineering optimization problem. Good problem formulation is the key to the success of an optimization study and is to a large degree an art. It is learned through practice and the study of successful applications and is based on the knowledge of the strengths, weaknesses, and peculiarities of the techniques provided by optimization theory. For these reasons, this text is liberally laced with engineering applications drawn from the literature and the experience of the authors. Moreover, along with presenting the techniques, we attempt to elucidate their relative advantages and disadvantages wherever possible by presenting or citing the results of actual computational tests. In the next several sections we discuss the elements of problem formulation in a bit more detail. In Section 1.2 we follow up this discussion by examining a few application formulations. 1.1.1

Defining the System Boundaries

Before undertaking any optimization study, it is important to clearly define the boundaries of the system under investigation. In this context a system is the restricted portion of the universe under consideration. The system boundaries are simply the limits that separate the system from the remainder of the universe. They serve to isolate the system from its surroundings, because, for purposes of analysis, all interactions between the system and its surroundings are assumed to be frozen at selected representative levels. Nonetheless, since interactions always exist, the act of defining the system boundaries is the first step in the process of approximating the real system. In many situations it may turn out that the initial choice of boundary is too restrictive. To fully analyze a given engineering system, it may be necessary to expand the system boundaries to include other subsystems that strongly affect the operation of the system under study. For instance, suppose a manufacturing operation has a paint shop in which finished parts are mounted on an assembly line and painted in different colors. In an initial study of the paint shop, we may consider it in isolation from the rest of the plant. However, we may find that the optimal batch size and color sequence we deduce for this system are strongly influenced by the operation of the fabrication department that produces the finished parts. A decision thus has

1.1

REQUIREMENTS FOR THE APPLICATION OF OPTIMIZATION METHODS

3

to be made whether to expand the system boundaries to include the fabrication system. An expansion of the system boundaries certainly increases the size and complexity of the composite system and thus may make the study much more difficult. Clearly, to make our work as engineers more manageable, we would prefer as much as possible to break down large complex systems into smaller subsystems that can be dealt with individually. However, we must recognize that such a decomposition may constitute a potentially misleading simplification of reality. 1.1.2

Performance Criterion

Given that we have selected the system of interest and have defined its boundaries, we next need to select a criterion on the basis of which the performance or design of the system can be evaluated so that the best design or set of operating conditions can be identified. In many engineering applications, an economic criterion is selected. However, there is a considerable choice in the precise definition of such a criterion: total capital cost, annual cost, annual net profit, return on investment, cost–benefit ratio, or net present worth. In other applications a criterion may involve some technological factors—for instance, minimum production time, maximum production rate, minimum energy utilization, maximum torque, maximum weight, and so on. Regardless of the criterion selected, in the context of optimization the best will always mean the candidate system with either the minimum or maximum value of the performance index. It is important to note that within the context of the optimization methods discussed in this book, only one criterion or performance measure can be used to define the optimum. It is not possible to find a solution that, say, simultaneously minimizes cost and maximizes reliability and minimizes energy utilization. This again is an important simplification of reality, because in many practical situations it would be desirable to achieve a solution that is best with respect to a number of different criteria. One way of treating multiple competing objectives is to select one criterion as primary and the remaining criteria as secondary. The primary criterion is then used as an optimization performance measure, while the secondary criteria are assigned acceptable minimum or maximum values and are treated as problem constraints. For instance, in the case of the paint shop study, the following criteria may well be selected by different groups in the company: 1. The shop foreman may seek a design that will involve long production runs with a minimum of color and part changes. This will maximize the number of parts painted per unit time. 2. The sales department would prefer a design that maximizes the inventory of parts of every type and color. This will minimize the time between customer order and order dispatch.

4


3. The company financial officer would prefer a design that will minimize inventories so as to reduce the amount of capital tied up in parts inventory. These are clearly conflicting performance criteria that cannot all be optimized simultaneously. A suitable compromise would be to select as the primary performance index the minimum annual cost but then to require as secondary conditions that the inventory of each part not be allowed to fall below or rise above agreed-upon limits and that production runs involve no more than some maximum acceptable number of part and color changes per week. In summary, for purposes of applying the methods discussed in this text, it is necessary to formulate the optimization problem with a single performance criterion. Advanced techniques do exist for treating certain types of multicriteria optimization problems. However, this new and growing body of techniques is quite beyond the scope of this book. The interested reader is directed to recent specialized texts [1, 2]. 1.1.3

Independent Variables

The third key element in formulating a problem for optimization is the selection of the independent variables that are adequate to characterize the possible candidate designs or operating conditions of the system. There are several factors to be considered in selecting the independent variables. First, it is necessary to distinguish between variables whose values are amenable to change and variables whose values are fixed by external factors, lying outside the boundaries selected for the system in question. For instance, in the case of the paint shop, the types of parts and the colors to be used are clearly fixed by product specifications or customer orders. These are specified system parameters. On the other hand, the order in which the colors are sequenced is, within constraints imposed by the types of parts available and inventory requirements, an independent variable that can be varied in establishing a production plan. Furthermore, it is important to differentiate between system parameters that can be treated as fixed and those that are subject to fluctuations influenced by external and uncontrollable factors. For instance, in the case of the paint shop, equipment breakdown and worker absenteeism may be sufficiently high to seriously influence the shop operations. Clearly, variations in these key system parameters must be taken into account in the formulation of the production planning problem if the resulting optimal plan is to be realistic and operable. Second, it is important to include in the formulation all the important variables that influence the operation of the system or affect the design definition. For instance, if in the design of a gas storage system we include the height, diameter, and wall thickness of a cylindrical tank as independent variables but exclude the possibility of using a compressor to raise the storage

1.1

REQUIREMENTS FOR THE APPLICATION OF OPTIMIZATION METHODS

5

pressure, we may well obtain a very poor design. For the selected fixed pressure, we would certainly find the least-cost tank dimensions. However, by including the storage pressure as an independent variable and adding the compressor cost to our performance criteria, we could obtain a design with a lower overall cost because of a reduction in the required tank volume. Thus, the independent variables must be selected so that all important alternatives are included in the formulation. In general, the exclusion of possible alternatives will lead to suboptimal solutions. Finally, another consideration in the selection of variables is the level of detail to which the system is considered. While it is important to treat all key independent variables, it is equally important not to obscure the problem by the inclusion of a large number of fine details of subordinate importance. For instance, in the preliminary design of a process involving a number of different pieces of equipment—pressure vessels, towers, pumps, compressors, and heat exchanges—one would normally not explicitly consider all the fine details of the design of each individual unit. A heat exchanger may well be characterized by a heat transfer surface area as well as shell-side and tubeside pressure drops. Detailed design variables such as number and size of tubes, number of tube and shell passes, baffle spacing, header type, and shell dimensions would normally be considered in a separate design study involving that unit by itself. In selecting the independent variables, a good rule is to include only those variables that have a significant impact on the composite system performance criterion. 1.1.4

System Model

Once the performance criterion and the independent variables have been selected, the next step in problem formulation is to assemble the model that describes the manner in which the problem variables are related and the way in which the performance criterion is influenced by the independent variables. In principle, optimization studies may be performed by experimenting directly with the system. Thus, the independent variables of the system or process may be set to selected values, the system operated under those conditions, and the system performance index evaluated using the observed performance. The optimization methodology would then be used to predict improved choices of the independent variable values and the experiments continued in this fashion. In practice, most optimization studies are carried out with the help of a simplified mathematical representation of the real system, called a model. Models are used because it is too expensive or time consuming or risky to use the real system to carry out the study. Models are typically used in engineering design because they offer the cheapest and fastest way of studying the effects of changes in key design variables on system performance. In general, the model will be composed of the basic material and energy balance equations, engineering design relations, and physical property equa-

6


tions that describe the physical phenomena taking place in the system. These equations will normally be supplemented by inequalities that define allowable operating ranges, specify minimum or maximum performance requirements, or set bounds on resource availabilities. In sum, the model consists of all elements that normally must be considered in calculating a design or in predicting the performance of an engineering system. Quite clearly, the assembly of a model is a very time consuming activity and one that requires a thorough understanding of the system being considered. In later chapters we will have occasion to discuss the mechanics of model development in more detail. For now, we simply observe that a model is a collection of equations and inequalities that define how the system variables are related and that constrain the variables to take on acceptable values. From the preceding discussion, we observe that a problem suitable for the application of optimization methodology consists of a performance measure, a set of independent variables, and a model relating the variables. Given these rather general and abstract requirements, it is evident that the methods of optimization can be applied to a very wide variety of applications. In fact, the methods we will discuss have been applied to problems that include the optimum design of process and structures, the planning of investment policies, the layout of warehouse networks, the determination of optimal trucking routes, the planning of heath care systems, the deployment of military forces, and the design of mechanical components, to name but a few. In this text our focus will be on engineering applications. Some of these applications and their formulations are discussed in the next section.

1.2

APPLICATIONS OF OPTIMIZATION IN ENGINEERING

Optimization theory finds ready application in all branches of engineering in four primary areas: 1. 2. 3. 4.

Design of components or entire systems Planning and analysis of existing operations Engineering analysis and data reduction Control of dynamic systems

In this section we briefly consider representative applications from each of the first three areas. The control of dynamic systems is an important area to which the methodology discussed in this book is applicable but which requires the consideration of specialized topics quite beyond the scope of this book. In considering the application of optimization methods in design and operations, keep in mind that the optimization step is but one step in the overall process of arriving at an optimal design or an efficient operation. Generally, that overall process will, as shown in Figure 1.1, consist of an iterative cycle

1.2


7

Figure 1.1. Engineering design process.

involving synthesis or definition of the structure of the system, model formulation, model parameter optimization, and analysis of the resulting solution. The final optimal design or new operating plan will be obtained only after solving a series of optimization problems, the solution to each of which will serve to generate new ideas for further system structures. In the interests of brevity, the examples in this section show only one pass of this iterative cycle and deal mainly with preparations for the optimization step. This focus should not be interpreted as an indication of the dominant role of optimization methods in the engineering design and systems analysis process. Optimization theory is a very powerful tool, but to be effective, it must be used skillfully and intelligently by an engineer who thoroughly understands the system under study. The primary objective of the following examples is simply to illustrate the wide variety but common form of the optimization problems that arise in the process of design and analysis.

8

1.2.1


Design Applications

Applications in engineering design range from the design of individual structural members to the design of separate pieces of equipment to the preliminary design of entire production facilities. For purposes of optimization, the shape or structure of the system is assumed to be known, and the optimization problem reduces to that of selecting values of the unit dimensions and operating variables that will yield the best value of the selected performance criterion. Example 1.1

Design of an Oxygen Supply System

Description. The basic oxygen furnace (BOF) used in the production of steel is a large fed-batch chemical reactor that employs pure oxygen. The furnace is operated in a cyclical fashion. Ore and flux are charged to the unit, treated for a specified time period, and then discharged. This cyclical operation gives rise to a cyclically varying demand rate for oxygen. As shown in Figure 1.2, over each cycle there is a time interval of length t1 of low demand rate D0 and a time interval t2 ⫺ t1 of high demand rate D1. The oxygen used in the BOF is produced in an oxygen plant in a standard process in which oxygen is separated from air by using a combination of refrigeration and distillation. Oxygen plants are highly automated and are designed to deliver oxygen at a fixed rate. To mesh the continuous oxygen plant with the cyclically operating BOF, a simple inventory system (Figure 1.3) consisting of a compressor and a storage tank must be designed. A number of design possibilities can be considered. In the simplest case, the oxygen plant capacity could be selected

Figure 1.2. Oxygen demand cycle, Example 1.1.

1.2


9

Figure 1.3. Design of oxygen production system, Example 1.1.

to be equal to D1, the high demand rate. During the low-demand interval the excess oxygen could just be vented to the air. At the other extreme, the oxygen plant capacity could be chosen to be just enough to produce the amount of oxygen required by the BOF over a cycle. During the low-demand interval, the excess oxygen produced would then be compressed and stored for use during the high-demand interval of the cycle. Intermediate designs could use some combination of venting and storage of oxygen. The problem is to select the optimal design. Formulation. The system of concern will consist of the O2 plant, the compressor, and the storage tank. The BOF and its demand cycle are assumed fixed by external factors. A reasonable performance index for the design is the total annual cost, which consists of the oxygen production cost (fixed and variable), the compressor operating cost, and the fixed costs of the compressor and storage vessel. The key independent variables are the oxygen plant production rate F (lb O2 /hr), the compressor and storage tank design capacities, H (HP) and V (ft3), respectively, and the maximum tank pressure p (psia). Presumably the oxygen plant design is standard so that the production rate fully characterizes the plant. Similarly, we assume that the storage tank will be of a standard design approved for O2 service. The model will consist of the basic design equations that relate the key independent variables. If Imax is the maximum amount of oxygen that must be stored, then using the corrected gas law we have V⫽

Imax RT z M p

(1.1)

10


where R T z M

⫽ ⫽ ⫽ ⫽

gas constant gas temperature (assume fixed) compressibility factor molecular weight of O2

From Figure 1.2, the maximum amount of oxygen that must be stored is equal to the area under the demand curve between t1 and t2 and D1 and F. Thus, Imax ⫽ (D1 ⫺ F)(t2 ⫺ t1)

(1.2)

Substituting (1.2) into (1.1), we obtain V⫽

(D1 ⫺ F)(t2 ⫺ t1) RT z M p

(1.3)

The compressor must be designed to handle a gas flow rate of (D1 ⫺ F)(t2 ⫺ t1)/t1 and to compress the gas to the maximum pressure p. Assuming isothermal ideal gas compression [3], H⫽

冉冊

p (D1 ⫺ F)(t2 ⫺ t1) RT ln t1 k1k2 p0

(1.4)

where k1 ⫽ unit conversion factor k2 ⫽ compressor efficiency p0 ⫽ O2 delivery pressure In addition to (1.3) and (1.4), the O2 plant rate F must be adequate to supply the total oxygen demand, or F⭓

D0t1 ⫹ D1(t2 ⫺ t1) t2

(1.5)

Moreover, the maximum tank pressure must be greater than the O2 delivery pressure, p ⭓ p0

(1.6)

The performance criterion will consist of the oxygen plant annual cost, C1 ($/yr) ⫽ a1 ⫹ a2F

(1.7)

where a1 and a2 are empirical constants for plants of this general type and include fuel, water, and labor costs. The capital cost of storage vessels is given by a power law correlation,

1.2


C2 ($) ⫽ b1V b2

11

(1.8a)

where b1 and b2 are empirical constants appropriate for vessels of a specific construction. The capital cost of compressors is similarly obtained from a correlation: C3 ($) ⫽ b3H b4

(1.8b)

The compressor power cost will, as an approximation, be given by b5t1H, where b5 is the cost of power. The total cost function will thus be of the form Annual cost ⫽ a1 ⫹ a2F ⫹ d(b1V b2 ⫹ b3H b4) ⫹ Nb5t1H

(1.9)

where N is the number of cycles per year and d is an appropriate annual cost factor. The complete design optimization problem thus consists of the problem of minimizing (1.9) by the appropriate choice of F, V, H, and p subject to Eqs. (1.3) and (1.4) as well as inequalities (1.5) and (1.6). The solution of this problem will clearly be affected by the choice of the cycle parameters (N, D0, D1, t1, and t2), the cost parameters (a1, a2, b1 –b5, and d), and the physical parameters (T, p0, k2, z, and M). In principle we could solve this problem by eliminating V and H from (1.9) using (1.3) and (1.4), thus obtaining a two-variable problem. We could then plot the contours of the cost function (1.9) in the plane of the two variables F and p, impose the inequalities (1.5) and (1.6), and determine the minimum point from the plot. However, the methods discussed in subsequent chapters allow us to obtain the solution with much less work. For further details and a study of solutions for various parameter values, the reader is invited to consult Jen et al. [4]. Example 1.1 presented a preliminary design problem formulation for a system consisting of several pieces of equipment. The next example illustrates a detailed design of a single structural element. Example 1.2

Design of a Welded Beam

Description. A beam A is to be welded to a rigid support member B. The welded beam is to consist of 1010 steel and is to support a force F of 6000 lb. The dimensions of the beam are to be selected so that the system cost is minimized. A schematic of the system is shown in Figure 1.4. Formulation. The appropriate system boundaries are quite self-evident. The system consists of the beam A and the weld required to secure it to B. The independent or design variables in this case are the dimensions h, l, t, and b, as shown in Figure 1.4. The length L is assumed to be specified at 14 in. For

12


Figure 1.4. Welded beam, Example 1.2.

notational convenience we redefine these four variables in terms of the vector of unknowns x: x ⫽ [x1, x2, x3, x4,]T ⫽ [h, l, t, b]T The performance index appropriate to this design is the cost of a weld assembly. The major cost components of such an assembly are (1) setup labor cost, (2) welding labor cost, and (3) material cost: F(x) ⫽ c0 ⫹ c1 ⫹ c2 where F(x) c0 c1 c2

⫽ ⫽ ⫽ ⫽

(1.10)

cost function setup cost welding labor cost material cost

Setup Cost c0. The company has chosen to make this component a weldment, because of the existence of a welding assembly line. Furthermore, assume that fixtures for setup and holding of the bar during welding are readily available. The cost c0 can therefore be ignored in this particular total-cost model. Welding Labor Cost c1. Assume that the welding will be done by machine at a total cost of $10/hr (including operating and maintenance expense). Furthermore, suppose that the machine can lay down a cubic inch of weld in 6 min. The labor cost is then

冉冊冉

c1 ⫽ 10

$ hr

冊冉冊

1 hr 60 min

6

冉冊

min $ V ⫽1 V in.3 w in.3 w

where Vw ⫽ weld volume, in.3 Material Cost c2 c2 ⫽ c3Vw ⫹ c4VB

1.2


13

where c3 ⫽ cost per volume of weld material, $/in.3, ⫽(0.37)(0.283) c4 ⫽ cost per volume of bar stock, $/in.3, ⫽(0.17)(0.283) VB ⫽ volume of bar A, in.3 From the geometry, Vw ⫽ 2(–21 h2l) ⫽ h2l and VB ⫽ tb(L ⫹ l) so c2 ⫽ c3h2l ⫹ c4tb(L ⫹ l) Therefore, the cost function becomes F(x) ⫽ h2l ⫹ c3h2l ⫹ c4tb(L ⫹ l)

(1.11)

or, in terms of the x variables, F(x) ⫽ (1 ⫹ c3)x21 x2 ⫹ c4 x3 x4(L ⫹ x2)

(1.12)

Not all combinations of x1, x2, x3, and x4 can be allowed if the structure is to support the load required. Several functional relationships between the design variables that delimit the region of feasibility must certainly be defined. These relationships, expressed in the form of inequalities, represent the design model. Let us first define the inequalities and then discuss their interpretation. The inequalities are g1(x) ⫽ ␶d ⫺ ␶ (x) ⭓ 0

(1.13)

g2(x) ⫽ ␴d ⫺ ␴ (x) ⭓ 0

(1.14)

g3(x) ⫽ x4 ⫺ x1 ⭓ 0

(1.15)

g4(x) ⫽ x2 ⭓ 0

(1.16)

g5(x) ⫽ x3 ⭓ 0

(1.17)

g6(x) ⫽ Pc(x) ⫺ F ⭓ 0

(1.18)

g7(x) ⫽ x1 ⫺ 0.125 ⭓ 0

(1.19)

g8(x) ⫽ 0.25 ⫺ ␦(x) ⭓ 0

(1.20)

14


where ␶d ␶ (x) ␴d ␴ (x) Pc(x) ␦(x)

⫽ ⫽ ⫽ ⫽ ⫽ ⫽

design shear stress of weld maximum shear stress in weld; a function of x design normal stress for beam material maximum normal stress in beam; a function of x bar buckling load; a function of x bar end deflection; a function of x

To complete the model, it is necessary to define the important stress states. Weld Stress ␶ (x). After Shigley [5], the weld shear stress has two components, ␶ ⬘ and ␶ ⴖ, where ␶ ⬘ is the primary stress acting over the weld throat area and ␶ ⴖ is a secondary torsional stress: ␶⬘ ⫽

F 兹2x1x2

␶ⴖ ⫽

and

MR J

with

冉冊冋冉冊册再冋冉

M⫽F L⫹ R⫽

x2 2

2

x22 x ⫹ x1 ⫹ 3 4 2

J ⫽ 2 0.707x1x2

1/2

x22 x ⫹ x1 ⫹ 3 12 2

冊册冎 2

where M ⫽ moment of F about center of gravity of weld group J ⫽ polar moment of inertia of weld group Therefore, the weld stress ␶ becomes ␶ (x) ⫽ [(␶ ⬘)2 ⫹ 2␶ ⬘␶ ⴖ cos ␪ ⫹ (␶ ⴖ)2]1 / 2

where cos ␪ ⫽

x2 2R

Bar Bending Stress ␴ (x). The maximum bending stress can be shown to be equal to ␴ (x) ⫽

6FL x4 x32

1.2


15

Bar Buckling Load Pc(x). If the ratio t/b ⫽ x3 /x4 grows large, there is a tendency for the bar to buckle. Those combinations of x3 and x4 that will cause this buckling to occur must be disallowed. It has been shown [6] that for narrow rectangular bars a good approximation to the buckling load is Pc(x) ⫽ where E I ␣ G

⫽ ⫽ ⫽ ⫽

冉

4.013兹EI␣ x 1⫺ 3 2 L 2L

冪␣冊 EI

Young’s modulus, ⫽30 ⫻ 106 psi 1 3 –– 12 x3 x4 1 –3 Gx3 x34 shearing modulus, ⫽12 ⫻ 106 psi

Bar Deflection ␦(x). To calculate the deflection, assume the bar to be a cantilever of length L. Thus, ␦(x) ⫽

4FL3 Ex33 x4

The remaining inequalities are interpreted as follows: Inequality g3 states that it is not practical to have the weld thickness greater than the bar thickness, Inequalities g4 and g5 are nonnegativity restrictions on x2 and x3. Note that the nonnegativity of x1 and x4 are implied by g3 and g7. Constraint g6 ensures that the buckling load is not exceeded. Inequality g7 specifies that it is not physically possible to produce an extremely small weld. Finally, the two parameters ␶d and ␴d in g1 and g2 depend on the material of construction. For 1010 steel, ␶d ⫽ 13,600 psi and ␴d ⫽ 30,000 psi are appropriate. The complete design optimization problem thus consists of the cost function (1.12) and the complex system of inequalities that results when the stress formulas are substituted into (1.13)–(1.20). All of these functions are expressed in terms of four independent variables. This problem is sufficiently complex that graphical solution is patently infeasible. However, the optimum design can readily be obtained numerically by using the methods of subsequent chapters. For a further discussion of this problem and its solution, see reference 7. 1.2.2

Operations and Planning Applications

The second major area of engineering application of optimization is found in the tuning of existing operations and development of production plans for multiproduct processes. Typically an operations analysis problem arises when an existing production facility designed under one set of conditions must be

16


adapted to operate under different conditions. The reasons for doing this might be as follows: 1. To accommodate increased production throughout 2. To adapt to different feedstocks or a different product slate 3. To modify the operations because the initial design is itself inadequate or unreliable The solution to such problems might require the selection of new temperature, pressure, or flow conditions; the addition of further equipment; or the definition of new operating procedures. Production planning applications arise from the need to schedule the joint production of several products in a given plant or to coordinate the production plans of a network of production facilities. Since in such applications the capital equipment is already in place, only the variable costs need to be considered. Thus, this type of application can often be formulated in terms of linear or nearly linear models. We will illustrate this class of applications using a refinery planning problem. Example 1.3

Refinery Production Planning

Description. A refinery processes crude oils to produce a number of raw gasoline intermediates that must subsequently be blended to make two grades of motor fuel, regular and premium. Each raw gasoline has a known performance rating, a maximum availability, and a fixed unit cost. The two motor fuels have a specified minimum performance rating and selling price, and their blending is achieved at a known unit cost. Contractual obligations impose minimum production requirements of both fuels. However, all excess fuel production or unused raw gasoline amounts can be sold in the open market at known prices. The optimal refinery production plan is to be determined over the next specified planning period. Formulation. The system in question consists of the raw gasoline intermediates, the blending operation, and the fluid motor fuels, as shown schematically in Figure 1.5. Excluded from consideration are the refinery processes involved in the production of the raw gasoline intermediates as well as the inventory and distribution subsystems for crudes, intermediates, and products. Since equipment required to carry out the blending operations is in place, only variable costs will be considered. The performance index in this case will be the net profit over the planning period. The net profit will be composed of motor fuel and intermediate sales minus blending costs minus the charged costs of the intermediates. The independent variables will simply be the flows depicted as directed arcs in Figure 1.5. Thus, each intermediate will have associated with it a variable that represents the amount of that intermediate allocated to the production of

1.2


17

Figure 1.5. Schematic of refinery planning problem, Example 1.3.

regular-grade gasoline, another that represents the amount used to make premium, and a third that represents the amount sold directly. Thus, for each intermediate i, xi ⫽ amount used for regular, bbl/period yi ⫽ amount used for premium, bbl/period zi ⫽ amount sold directly, bbl/period Each product will have two variables associated with it: one to represent the contracted sales and one to represent the open-market sales. Thus, for each product j, uj ⫽ amount allocated to contracts, bbl/period vj ⫽ amount sold in open market, bbl/period

The model will consist of material balances on each intermediate and product, blending constraints that ensure that product performance ratings are met, and bounds on the contract sales:

18


1. Material balances on each intermediate i: xi ⫹ yi ⫹ zi ⭐ ␣i

(1.21)

where ␣i is the availability of intermediate i over the period, in bbl/ period. 2. Material balances on each product:

冘x ⫽u ⫹v i

1

冘y ⫽u ⫹v

1

i

i

2

(1.22)

2

i

3. Blending constraints on each product:

冘 ␤ x ⭓ ␥ (u ⫹ v ) 冘 ␤ y ⭓ ␥ (u ⫹ v ) i i

1

1

1

i i

i

2

2

(1.23)

2

i

where ␤i is the performance rating of intermediate i and ␥j is the minimum performance rating of product j. 4. Contract sales restrictions for each product j. uj ⭓ ␦j

(1.24)

where ␦j is the minimum contracted production, in bbl/period. The performance criterion (net profit) is given by

冘c

u ⫹

(1) j j

where c(1) i c(2) j c(3) i c(4) i c(5) i

⫽ ⫽ ⫽ ⫽ ⫽

冘c

v ⫹

(2) j j

冘c

z ⫺

(3) i i

冘c

(xi ⫹ yi ⫹ zi) ⫺

(4) i

i

冘c

(xi ⫹ yi)

(5) i

i

unit selling price for contract sales of j unit selling price for open-market sales of j unit selling price of direct sales of intermediate i unit charged cost of intermediate i blending cost of intermediate i

Using the data given in Table 1.1, the planning problem reduces to Maximize 40u1 ⫹ 55u2 ⫹ 46v1 ⫹ 60v2 ⫹ 6z1 ⫹ 8z2 ⫹ 7.50z3 ⫹ 7.50z4 ⫹ 20z5 ⫺ 25(x1 ⫹ y1) ⫺ 28(x2 ⫹ y2) ⫺ 29.50(x3 ⫹ y3) ⫺ 35.50(x4 ⫹ y4) ⫺ 41.50(x5 ⫹ y5)

Subject to Constraints of type (1.21):

1.2

Table 1.1


19

Data for Example 1.3

Raw Gasoline Intermediate 1 2 3 4 5

Product Type Regular (1) Premium (2)

Availability, ␣i (bbl / period) 2 4 4 5 5

⫻ ⫻ ⫻ ⫻ ⫻

Performance Rating, ␤i

Selling Price, c(3) i

Charged Cost, c(4) i

Blending Cost, c(5) i

70 80 85 90 99

30.00 35.00 36.00 42.00 60.00

24.00 27.00 28.50 34.50 40.00

1.00 1.00 1.00 1.00 1.50

105 105 105 105 105

Selling Price ($ / bbl)

Minimum Contract Sales ␦j

Minimum Performance Rating

Contract, c(1) j

Open Market, c(2) j

5 ⫻ 105 4 ⫻ 105

85 95

$40.00 $55.00

$46.00 $60.00

x1 ⫹ y1 ⫹ z1 ⭐ 2 ⫻ 105 x2 ⫹ y2 ⫹ z2 ⭐ 4 ⫻ 105 x3 ⫹ y3 ⫹ z3 ⭐ 4 ⫻ 105 x4 ⫹ y4 ⫹ z4 ⭐ 5 ⫻ 105 x5 ⫹ y5 ⫹ z5 ⭐ 5 ⫻ 105 Constraints of type (1.22): x1 ⫹ x2 ⫹ x3 ⫹ x4 ⫹ x5 ⫽ u1 ⫹ v1 y1 ⫹ y2 ⫹ y3 ⫹ y4 ⫹ y5 ⫽ u2 ⫹ v2 Constraints of type (1.23): 70x1 ⫹ 80x2 ⫹ 85x3 ⫹ 90x4 ⫹ 99x5 ⭓ 85(u1 ⫹ v1) 70y1 ⫹ 80y2 ⫹ 85y3 ⫹ 90y4 ⫹ 99y5 ⭓ 95(u2 ⫹ v2) Constraints of type (1.24): u1 ⭓ 5 ⫻ 105

u2 ⭓ 4 ⫻ 105

In addition, all variables must be greater than or equal to zero to be physically realizable. The composite optimization problem involves 19 vari-

20


ables and 11 constraints plus the nonnegativity conditions. Note that all model functions are linear in the independent variables. In general, refineries will involve many more intermediate streams and products than were considered in this example. Moreover, in practice it may be important to include further variables, reflecting the material in inventory, as well as to expand the model to cover several consecutive planning periods. In the latter case, a second subscript could be added to the set of variables, for example, xik ⫽ amount of intermediate i used for regular grade in planning period k The resulting production planning model can then become very large. In practice, models of this type with over a thousand variables are solved quite routinely. 1.2.3

Analysis and Data Reduction Applications

A further fertile area for the application of optimization techniques in engineering can be found in nonlinear regression problems as well as in many analysis problems arising in engineering science. A very common problem arising in engineering model development is the need to determine the parameters of some semitheoretical model given a set of experimental data. This data reduction or regression problem inherently transforms to an optimization problem, because the model parameters must be selected so that the model fits the data as closely as possible. Suppose some variable y assumed to be dependent upon an independent variable x and related to x through a postulated equation y ⫽ ƒ(x, ␪1, ␪2) that depends upon two parameters ␪1 and ␪2. To establish the appropriate values of ␪1 and ␪2, we run a series of experiments in which we adjust the independent variable x and measure the resulting y. As a result of a series of N experiments covering the range of x of interest, a set of y and x values (yi, xi), i ⫽ 1, . . . , N, is available. Using these data, we now try to ‘‘fit’’ our function to the data by adjusting ␪1 and ␪2 until we get a ‘‘good fit.’’ The most commonly used measure of a good fit is the least-squares criterion,

冘 [y ⫺ ƒ(x , ␪ , ␪ )] N

L(␪i, ␪2) ⫽

2

i

i

1

2

(1.25)

i⫽1

The difference yi ⫺ ƒ(xi, ␪1, ␪2) between the experimental value yi and the predicted value ƒ(xi, ␪1, ␪2) measures how close our model prediction is to the data and is called the residual. The sum of the squares of the residuals at all the experimental points gives an indication of goodness of fit. Clearly, if L(␪1, ␪2) is equal to zero, then the choice of ␪1, ␪2 has led to a perfect fit; the

1.2


21

data points fall exactly on the predicted curve. The data-fitting problem can thus be viewed as an optimization problem in which L(␪1, ␪2) is minimized by appropriate choice of ␪1 and ␪2. Example 1.4

Nonlinear Curve Fitting

Description. The pressure–molar volume–temperature relationship of real gases is known to deviate from that predicted by the ideal gas relationship, Pv ⫽ RT where P ⫽ pressure, atm v ⫽ molar volume, cm2 /g 䡠 mol T ⫽ temperature, K R ⫽ gas constant, 82.06 atm 䡠 cm3 /g 䡠 mol 䡠 K The semiempirical Redlich–Kwong equation [3] P⫽

RT v⫺b

⫺

a T 1 / 2v(v ⫹ b)

(1.26)

is intended to correct for the departure from ideality, but it involves two empirical constants a and b whose values are best determined from experimental data. A series of PvT measurements, listed in Table 1.2, are made for CO2, from which a and b are to be estimated using nonlinear regression. Formulation. Parameters a and b will be determined by minimizing the leastsquares function (1.25). In the present case, the function will take the form

冘冋P ⫺ v RT⫺ b ⫹ T 8

i

i

i⫽1

Table 1.2

i

册

a 1/2 i vi (vi ⫹ b)

2

(1.27)

PVT Data for CO2

Experiment Number

P, atm

v, cm3 / g 䡠 mol

T, K

1 2 3 4 5 6 7 8

33 43 45 26 37 39 38 63.6

500 500 600 700 600 700 400 400

273 323 373 273 323 373 273 373

22


where Pi is the experimental value at experiment i and the remaining two terms correspond to the value of P predicted from Eq. (1.26) for the conditions of experiment i for some selected values of the parameters a and b. For instance, the term corresponding to the first experimental point will be

冉

33 ⫺

冊

82.06(273) a ⫹ 1/2 500 ⫺ b (273) (500)(500 ⫹ b)

2

Function (1.27) is thus a two-variable function whose value is to be minimized by appropriate choice of the independent variables a and b. If the Redlich–Kwong equation were to precisely match the data, then at the optimum the function (1.27) would be exactly equal to zero. In general, because of experimental error and because the equation is too simple to accurately model the CO2 nonidealities, Eq. (1.27) will not be equal to zero at the optimum. For instance, the optimal values of a ⫽ 6.377 ⫻ 107 and b ⫽ 29.7 still yield a squared residual of 9.7 ⫻ 10⫺2. In addition to regression applications, a number of problems arise in engineering science that can be solved by posing them as optimization problems. One rather classical application is the determination of the equilibrium composition of a chemical mixture [3]. It is known that the equilibrium composition of a closed system at fixed temperature and pressure with specified initial composition will be that composition that minimizes the Gibbs free energy of the system. As shown by White et al. [8], the determination of the equilibrium composition can thus be posed as the problem of minimizing a nonlinear function subject to a set of linear equations in nonnegative variables. Another classical engineering problem that can be posed and solved as an optimization problem is the determination of the steady-state current flows in an electrical resistance network [9]. Given a network with specified arc resistances and a specified overall current flow, the arc current flows can be determined as the solution of the problem of minimizing the total I 2R power loss subject to a set of linear constraints that ensure that Kirchhoff’s current law is satisfied at each arc junction in the network. Example 1.5

Metal Cutting

One of the engineering applications of mathematical programming is the problem of determining the optimal machining parameters in metal cutting. A detailed discussion of different optimization models in metal cutting with illustrations is given in a survey paper by Philipson and Ravindran [10]. Here we shall discuss a machining problem in which a single cutting tool turns a diameter in one pass. The decision variables in this machining problem are the cutting speed v and the feed per revolution ƒ. Increasing the speed and feed reduces the actual

1.2

23


machining time and hence the machining cost, but it has an adverse effect on the life of the cutting tool and results in a higher tooling cost. In addition, the optimal values of v and ƒ will also depend on labor and overhead costs of nonproductive time and tool-changing time. Hence, minimization of the total cost per component is the criterion most often used in selecting the optimal machining parameters, feed and speed. The cost per component c for a part produced in one pass is given by Armarego and Brown [11]: c ⫽ (cost of nonproductive time/component) ⫹ (machining time cost) ⫹ (cost of tool-changing time/component) ⫹ (tool cost/component)

Material costs are not considered. The third and fourth terms may be stated more specifically as Cost of tool-changing time per component ⫽

cost rate ⫻ tool-changing time number of parts produced between tool changes

Tool cost per component ⫽

tool cost per cutting edge number of parts produced between tool changes

and

The cost equation can be expressed mathematically as

冉冊冉冊

c ⫽ xTL ⫹ xTc ⫹ xTd where x TL Tc Tac T Td y T/Tac

⫽ ⫽ ⫽ ⫽ ⫽ ⫽ ⫽ ⫽

Tac T ⫹ y ac T T

(dollars)

(1.28)

labor plus overhead cost rate, $ nonproductive time (loading, unloading, and inspection time), min machining time, including approach time, min actual cutting time (approximately equal to Tc), min tool life, min [given by Eq. (1.30)] tool-changing time, min tool cost per cutting edge, $ number of parts produced between tool changes

The equation for machining time Tc is Tc ⫽

l ␭vƒ

(min)

(1.29)

24


where l ⫽ distance traveled by the tool in making a turning pass, in. ␭ ⫽ 12/ ␲D, where D is the mean workpiece diameter, in. v ⫽ cutting speed, surface feet/min ƒ ⫽ feed, in./rev It has been found [11] that tool life, cutting speed, and feed are related as follows: T⫽

A v1 / nƒ1 / n1

(min)

(1.30)

where A, n, and n1 are constants. Assuming that Tac ⯝ Tc and inserting Eqs. (1.29) and (1.30) into Eq. (1.28), we obtain c ⫽ xTL ⫹

冉

冊

xl l yl (1 / n)⫺1 (1 / n1)⫺1 ⫹ xTd ⫹ v ƒ ␭vƒ ␭A ␭A

(dollars)

(1.31)

The constraints imposed on v and ƒ by the machine tool and by process and part requirements are as follows: (i) Maximum and Minimum Available Cutting Speed vmin ⭐ v ⭐ vmax

(ii) Maximum and Minimum Available Feed ƒmin ⭐ ƒ ⭐ ƒmax (iii) Maximum Allowable Cutting Force (Ft,max). This constraint is necessary to limit tool work deflections and their effect upon the accuracy of the turned part. Armarego and Brown [11] give the following expression for the tangential cutting force Ft: Ft ⫽ cf ƒ␣dc␥

(1.32)

where ct, ␣, and ␥ are constants and dc is the depth of cut, which is held constant at a given value. This constraint on the cutting force results in the following feed constraint:

冋册

Ft,max ƒ⭐ ct d␥c

1/␣

(iv) Maximum Available Horsepower. The horsepower consumed in cutting can be determined from the equation

1.2


HP ⫽

25

Ft v 33,000

where Ft is obtained from Eq. (1.32). If Pmax is the maximum horsepower available at the spindle, then vƒ␣ ⭐

Pmax(33,000) ct d␥c

For a given Pmax, ct, ␥, and dc, the right-hand side of the above inequality will be a constant. (v) Stable Cutting Region. Certain combinations of v and ƒ values are likely to cause chatter vibration, adhesion, and built-up edge formation. To avoid this problem, the following stable cutting region constraint is used: v␦ƒ ⭓ ␤

where ␦ and ␤ are given constants. As an illustration [10], consider the case where a single diameter is to be turned in one pass using the feed rate and cutting speed that will minimize costs. The bar is 2.75 in. in diameter by 12.00 in. long. The turned bar is 2.25 in. in diameter by 10.00 in. long. In the cutting speed calculations, a mean diameter of 2.50 in. will be used. The lathe has a 15-HP motor and a maximum speed capability of 1500 rpm. The minimum speed available is 75 rpm. The cost rate, tool costs, ideal time, tool-changing time, and tool life parameters are given below: x ⫽ $0.15/min

TL ⫽ 2.00 min

l ⫽ 10.00 in.

D ⫽ 2.50 in.

␭ ⫽ 1.528

Td ⫽ 1.00 min

y ⫽ $0.50

A ⫽ 113,420

n ⫽ 0.30

n1 ⫽ 0.45

dc ⫽ 0.25 in. Nmax ⫽ 1500 rpm Ft,max ⫽ 1583.0 lb ␣ ⫽ 0.78 ␦ ⫽ 2.0

Nmin ⫽ 75 rpm ct ⫽ 344.7 ␥ ⫽ 0.9

Machine drive efficiency ⫽ 0.8 Machine HP ⫽ 15.0

␤ ⫽ 380,000

When the fixed values are inserted into Eq. (1.28), the cost function becomes

26


Minimize c ⫽ 0.30 ⫹

982.0 ⫹ 8.1 ⫻ 10⫺9v2.333ƒ1.222 vƒ

(Note: For ease of calculations, ƒ is expressed in thousandths of an inch per revolution rather than in inches per revolution.) The constraints on v and ƒ are given by v ⭐ 982.0 v ⭓ 49.1

ƒ ⭐ 35.0

(cutting force)

ƒ ⭓ 1.0 vƒ 0.78 ⭐ 4000.0 v ƒ ⭓ 380,000.0 2.0

(horsepower) (stable cutting region)

v, ƒ ⭓ 0 1.2.4

Classical Mechanics Applications

The methods described in this book can be useful in a number of application areas, including engineering design, and various forms of data approximation. A student of this field learns that design is the essence of engineering and optimization is the essence of design. On the other hand, it is now clear that the optimization philosophy given in this text is also applicable to problems in classical mechanics and many other fields not considered by the authors. Ragsdell and Carter proposed the ‘‘energy method’’ [12] as an alternative to more traditional approaches to the solution of conservative, linear and nonlinear initial-value problems. The energy method employs a minimum principle and avoids most of the difficulties associated with the solution of nonlinear differential equations by formulation and iterative solution of an appropriate sequence of optimization problems. It is interesting to note that solutions to almost any predetermined accuracy level are possible and that the energy method appears to be especially useful for very nonlinear problems with solutions that progress over long periods of time and/or space. The method has been used to predict the position of spacecraft over long distances and the position of the robot arm on the Space Shuttle. Carter and Ragsdell [13] have given a direct solution to the optimal column problem of Lagrange. This problem is of interest because it defied solution for so long. The problem statement is simple: Given an amount of material with known physical properties (e.g., 1010 mild steel) that is to be formed into a column and pinned at each end, what is the optimal tapering function if we assume uniform, solid cross sections? (See Figure 1.6.) Lagrange worked on this problem most of his life and published [14] his conclusions shortly before his death. He was led to conclude that the optimal

1.2


27

D (x)

Pc

Pc

0

X

l Figure 1.6. Optimal column.

column was a right circular cylinder. Lagrange suspected and many later proved that this could not be so. Carter and Ragsdell later showed that a direct optimization approach employing Fourier approximating functions would produce a valid solution to the optimal column problem of Lagrange and to many related minimum-weight structural design problems. The reader is referred to the references for additional details. 1.2.5

Taguchi System of Quality Engineering

The American, British, and European schools of thought on optimization and optimal design have developed along rather different paths than the work in Japan since the early work of George Dantzig [14] and others during and after World War II. The American, British, and European approach is essentially monolithic, whereas there are some important differences to be noted in the Japanese approach. It is not our purpose to judge whether one path is best, but simply to note that they appear to be significantly different. In the United States many optimization methods have been developed and used based on the assumption that an analytical (mathematical) model or simulation is available. This assumption is used much less in Japan. On the contrary, the common assumption seen is that the system performance will be available through direct measurement of the system being optimized. In addition, in Japan there seems to be much greater emphasis on stochastic formulations and a desire to understand the resulting variation propagation. Genichi Taguchi is the father of what is now called quality engineering in Japan. He was asked to help with the poor telephone system performance in Japan after the war. It was difficult to have a phone conversation lasting more than a few minutes. It would have been easy to suggest higher quality components to achieve better reliability, but funding was poor, so he took another approach. Taguchi asked, ‘‘Is it possible to achieve good performance with inexpensive components?’’ He wondered, ‘‘Can we control the propagation of unwanted variation?’’ This led him to develop product parameter design, which is now known as the first of four actions in the Taguchi system of quality engineering (Figure 1.7).

28


Figure 1.7. Taguchi’s system of quality engineering.

Taguchi’s approach is based on an additive model or linear independence of factor effects. He seeks to find the control factor levels, which will attenuate variation propagation, so that the product will continue to perform at target even in the presence of internal and external variation. The emphasis is on finding better, more stable designs, not necessarily optimal designs. The approach employs direct sampling of system performance typically using prototypes and very compact sampling strategies. Orthogonal arrays are used to estimate system sensitivities. Taguchi’s work is beyond the scope of this book, but it is important work and the interested reader is advised to consult the many excellent references [15–18].

1.3

STRUCTURE OF OPTIMIZATION PROBLEMS

Although the application problems discussed in the previous section originate from radically different sources and involve different systems, at their root they have a remarkably similar form. All four can be expressed as problems requiring the minimization of a real-valued function ƒ(x) of an N-component vector argument x ⫽ (x1, x2, . . . , xN) whose values are restricted to satisfy a number of real-valued equations hk(x) ⫽ 0, a set of inequalities gj (x) ⭓ 0, ⭓ xi ⭓ x (L) and the variable bounds x (U) i i . In subsequent discussions we will refer to the function ƒ(x) as the objective function, to the equations hk(x) ⫽ 0 as the equality constraints, and to the inequalities gj (x) ⭓ 0 as the inequality constraints. For our purposes, these problem functions will always be assumed to be real valued, and their number will always be finite. The general problem

1.4

Minimize

SCOPE OF THIS BOOK

29

ƒ(x)

Subject to hk(x) ⫽ 0

k ⫽ 1, . . . , K

gj (x) ⭓ 0 x

(U) i

⭓ xi ⭓ x

j ⫽ 1, . . . , J (L) i

i ⫽ 1, . . . , N

is called the constrained optimization problem. For instance, Examples 1.1, 1.2, and 1.3 are all constrained problems. The problem in which there are no constraints, that is, J⫽K⫽0 and x (U) ⫽ ⫺x (L) ⫽⬁ i i

i ⫽ 1, . . . , N

is called the unconstrained optimization problem. Example 1.4 is an unconstrained problem. Optimization problems can be classified further based on the structure of the functions ƒ, hk, and gj and on the dimensionality of x. Unconstrained problems in which x is a one-component vector are called single-variable problems and form the simplest but nonetheless very important subclass. Constrained problems in which the function hk and gj are all linear are called linearly constrained problems. This subclass can further be subdivided into those with a linear objective function ƒ and those in which ƒ is nonlinear. The category in which all problem functions are linear in x includes problems with continuous variables, which are called linear programs, and problems in integer variables, which are called integer programs. Example 1.3 is a linear programming problem. Problems with nonlinear objective and linear constraints are sometimes called linearly constrained nonlinear programs. This class can further be subdivided according to the particular structure of the nonlinear objective function. If ƒ(x) is quadratic, the problem is a quadratic program; if it is a ratio of linear functions, it is called a fractional linear program; and so on. Subdivision into these various classes is worthwhile because the special structure of these problems can be efficiently exploited in devising solution techniques. We will consider techniques applicable to most of these problem structures in subsequent chapters.

1.4

SCOPE OF THIS BOOK

In this text we study the methodology applicable to constrained and unconstrained optimization problems. Our primary focus is on general-purpose tech-

30


niques applicable to problems in continuous variables, involving real-valued constraint functions and a single real-valued objective function. Problems posed in terms of integer or discrete variables are considered only briefly. Moreover, we exclude from consideration optimization problems involving functional equations, non-steady-state models, or stochastic elements. These very interesting but more complex elements are the appropriate subject matter for more advanced, specialized texts. While we make every effort to be precise in stating mathematical results, we do not normally present detailed proofs of these results unless such a proof serves to explain subsequent algorithmic constructions. Generally we simply cite the original literature source for proofs and use the pages of this book to motivate and explain the key concepts underlying the mathematical constructions. One of the goals of this text is to demonstrate the applicability of optimization methodology to engineering problems. Hence, a considerable portion is devoted to engineering examples, to the discussion of formulation alternatives, and to consideration of computational devices that expedite the solution of applications problems. In addition, we review and evaluate available computational evidence to help elucidate why particular methods are preferred under certain conditions. In Chapter 2, we begin with a discussion of the simplest problem, the single-variable unconstrained problem. This is followed by an extensive treatment of the multivariable unconstrained case. In Chapter 4, the important linear programming problem is analyzed. With Chapter 5 we initiate the study of nonlinear constrained optimization by considering tests for determining optimality. Chapters 6–10 focus on solution methods for constrained problems. Chapter 6 considers strategies for transforming constrained problems into unconstrained problems, while Chapter 7 discusses direct-search methods. Chapters 8 and 9 develop the important linearization-based techniques, and Chapter 10 discusses methods based on quadratic approximations. Then, in Chapter 11 we summarize some of the methods available for specially structured problems. Next, in Chapter 12 we review the results of available comparative computational studies. The text concludes with a survey of strategies for executing optimization studies (Chapter 13) and a discussion of three engineering case studies (Chapter 14). REFERENCES 1. Zeleny, M., Multiple Criteria Decision Making, McGraw-Hill, New York, 1982. 2. Vincent, T. L., and W. J. Grantham, Optimality in Parametric Systems, Wiley, New York, 1981. 3. Bett, K. E., J. S. Rowlinson, and G. Saville, Thermodynamics for Chemical Engineers, MIT Pres, Cambridge, MA, 1975. 4. Jen, F. C., C. C. Pegels, and T. M. Dupuis, ‘‘Optimal Capacities of Production Facilities,’’ Manag. Sci. 14B, 570–580 (1968).

REFERENCES

31

5. Shigley, J. E., Mechanical Engineering Design, McGraw-Hill, New York, 1973, p. 271. 6. Timoshenko, S., and J. Gere, Theory of Elastic Stability, McGraw-Hill, New York, 1961, p. 257. 7. Ragsdell, K. M., and D. T. Phillips, ‘‘Optimal Design of a Class of Welded Structures using Geometric Programming,’’ ASME J. Eng. Ind. Ser. B, 98(3), 1021– 1025 (1975). 8. White, W. B., S. M. Johnson, and G. B. Dantzig, ‘‘Chemical Equilibrium in Complex Mixtures,’’ J. Chem. Phys., 28, 251–255 (1959). 9. Hayt, W. H., and J. E. Kemmerly, Engineering Circuit Analysis, McGraw-Hill, New York, 1971, Chap. 2. 10. Philipson, R. H., and A. Ravindran, ‘‘Application of Mathematical Programming to Metal Cutting,’’ Math. Prog. Study, 11, 116–134 (1979). 11. Armarego, E. J. A., and R. H. Brown, The Machining of Metals, Prentice-Hall, Englewood Cliffs, NJ, 1969. 12. Ragsdell, K. M., and W. J. Carter, ‘‘The Energy Method,’’ Mechanisms and Machine Theory, 10(3), 245–260 (1975). 13. Carter, W. J., and K. M. Ragsdell, ‘‘The Optimal Column,’’ ASME J. Eng. Mater. Technol., 71–76 (1974). 14. Trefethen, F. N., ‘‘A History of Operations Research,’’ in Operations Research for Management (J. F. McCloskey and F. N. Trefethen, Eds.), John Hopkins Press, Baltimore, MD, 1954. 15. Ealey, L. A., Quality by Design, ASI Press, Dearborn, MI, 1988. 16. Taguchi, G. Taguchi on Robust Technology Development, ASME Press, New York, 1993. 17. Taguchi, G., System of Experimental Design, Vols. 1 and 2, Unipub / Kraus, White Plains, NY, 1987. 18. Taguchi, G., Introduction to Quality Engineering, Asian Productivity Organization, Tokyo, 1986.

2 FUNCTIONS OF A SINGLE VARIABLE The optimization problem in which the performance measure is a function of one variable is the most elementary type of optimization problem. Yet, it is of central importance to optimization theory and practice, not only because it is a type of problem that the engineer commonly encounters in practice, but also because single-variable optimization often arises as a subproblem within the iterative procedures for solving multivariable optimization problems. Because of the central importance of the single-variable problem, it is not surprising that a large number of algorithms have been devised for its solution. Basically, these methods can be categorized according to the nature of the function ƒ(x) and the variable x involved as well as the type of information that is available about ƒ(x).

2.1

PROPERTIES OF SINGLE-VARIABLE FUNCTIONS

A function ƒ(x), in the most elementary sense, is a rule that assigns to every choice of x a unique value y ⫽ ƒ(x). In this case, x is called the independent variable and y is called the dependent variable. Mathematically, consider a set S 傺 R, where R is the set of all real numbers. We can define a correspondence or a transformation that assigns a single numerical value to every point x 僆 S. Such a correspondence is called a scalar function ƒ defined on the set S. When the set S ⫽ R, we have an unconstrained function of one variable. When S is a proper subset of R, we have a function defined on a constrained region. For example,

32


2.1

PROPERTIES OF SINGLE-VARIABLE FUNCTIONS

ƒ(x) ⫽ x 3 ⫹ 2x 2 ⫺ x ⫹ 3

33

for all x 僆 R

is an unconstrained function, while ƒ(x) ⫽ x 3 ⫹ 2x 2 ⫺ x ⫹ 3

for all x 僆 S ⫽ {x兩 ⫺ 5 ⭐ x ⭐ 5}

is a constrained function. In engineering optimization, we call ƒ the objective function and S the feasible region, the constraint set, or the domain of interest of x. Most physical processes can be modeled or described by continuous functions, that is, functions that are continuous at every point x in their domain. However, it is not at all unusual in engineering applications to encounter discontinuous functions. For instance, if we construct a function that measures the cost of heating per Btu at different delivery temperatures, we quite naturally arrive at the broken curve shown in Figure 2.1. The cost is a discontinuous function of the delivery temperature; however, the temperature itself can assume all values over the range of, say, 200–3000⬚F. Of course, it is not necessary that the domain of the independent variable x assume all real values over the range in question. It is quite possible that the variable could assume only discrete values. For instance, if we construct a function that gives the cost of commercial pipe per foot of length at different pipe diameters, we quite naturally arrive at the sequence of discrete points shown in Figure 2.2. Pipe is made only in a finite number of sizes. Note: It is important to remember the following properties of continuous functions: 1. A sum or product of a continuous function is continuous. 2. The ratio of two continuous functions is continuous at all points where the denominator does not vanish. Clearly, depending upon whether the function to be optimized is continuous or discontinuous or depending upon the nature of its domain, different meth-

Figure 2.1. Discontinuous function.

34

FUNCTIONS OF A SINGLE VARIABLE

Figure 2.2. Discrete function.

ods will have to be employed to conduct the search for the optimum. A method that is effective for continuous functions will probably not be effective for discontinuous functions, although the converse may be possible. In addition to these properties, functions can also be classified according to their shape or topology within the range of interest. Monotonic Functions. A function ƒ(x) is monotonic (either increasing or decreasing) if for any two points x1 and x2, with x1 ⭐ x2, it follows that

ƒ(x1) ⭐ ƒ(x2)

(monotonically increasing)

ƒ(x1) ⭓ ƒ(x2)

(monotonically decreasing)

Figure 2.3 illustrates a monotonically increasing function, and Figure 2.4 illustrates a monotonically decreasing function. Note that a function does not have to be continuous to be monotonic. Figure 2.5 illustrates a function that is monotonically decreasing for x ⭐ 0 and increasing for x ⭓ 0. The function attains its minimum at x ⫽ x* (origin), and it is monotonic on either side of the minimum point. Such a function is called a unimodal function. Definition A function ƒ(x) is unimodal on the interval a ⭐ x ⭐ b if and only if it is monotonic on either side of the single optimal point x* in the interval. In

Figure 2.3. Monotonic increasing function.

2.2

OPTIMALITY CRITERIA

35

Figure 2.4. Monotonic decreasing function.

other words, if x* is the single minimum point of ƒ(x) in the range a ⭐ x ⭐ b, then ƒ(x) is unimodal on the interval if and only if for any two points x1 and x2 x* ⭐ x1 ⭐ x2 implies that ƒ(x*) ⭐ ƒ(x1) ⭐ ƒ(x2) and x* ⭓ x1 ⭓ x2 implies that ƒ(x*) ⭐ ƒ(x1) ⭐ ƒ(x2) As shown in Figure 2.6, a function does not have to be continuous to be unimodal. Unimodality is an extremely important functional property used in optimization. We shall discuss its usefulness in Section 2.3.

2.2

OPTIMALITY CRITERIA

In considering optimization problems, two questions generally must be addressed: 1. Static Question. How can one determine whether a given point x* is the optimal solution? 2. Dynamic Question. If x* is not the optimal point, then how does one go about finding a solution that is optimal?

Figure 2.5. Unimodal function.

36


Figure 2.6. Unimodal functions.

In this section we are concerned primarily with the static question, namely, developing a set of optimality criteria for determining whether a given solution is optimal. Definitions A function ƒ(x) defined on a set S attains its global minimum at a point x** 僆 S if and only if ƒ(x**) ⭐ ƒ(x)

for all x 僆 S

A function ƒ(x) defined on S has a local minimum (relative minimum) at a point x* 僆 S if and only if ƒ(x*) ⭐ ƒ(x)

for all x within a distance ␧ from x*

that is, there exists an ␧ ⬎ 0 such that, for all x satisfying 兩x ⫺ x*兩 ⬍ ␧, ƒ(x*) ⭐ ƒ(x). Remarks 1. By reversing the directions of inequality, we can get the equivalent definitions of global maximum and local maximum. 2. Under the assumption of unimodality, the local minimum automatically becomes the global minimum. 3. When the function is not unimodal, multiple local optima are possible and the global minimum can be found only by locating all local optima and selecting the best one. In Figure 2.7, x1 is the global maximum, x2 is a local minimum, x3 is a local maximum, x4 is the global minimum, and x5 may be considered as both local minimum and local maximum points. Identification of Single-Variable Optima. Suppose ƒ(x) is a function of a single variable x defined on an open interval (a, b) and ƒ is differentiable to the nth order over the interval. If x* is a point within that interval, then Taylor’s

2.2

37

OPTIMALITY CRITERIA

Figure 2.7. Local and global optima.

theorem allows us to write the change in the value of ƒ from x* to x* ⫹ ␧ as follows: ƒ(x* ⫹ ␧) ⫺ ƒ(x*) ⫽ (␧)

冏

dƒ dx

⫹

x⫽x*

␧2 d 2ƒ

2! dx 2

冏

⫹䡠䡠䡠⫹

x⫽x*

␧n dnƒ

n! dxn

冏

x⫽x*

⫹ On⫹1(␧)

(2.1)

where On⫹1(␧) indicates terms of (n ⫹ 1)st order or higher in ␧. If x* is a local minimum of ƒ on (a, b), then from the definition there must be an ␧ neighborhood of x* such that for all x within a distance ␧ ƒ(x) ⭓ ƒ(x*)

(2.2)

Inequality (2.2) implies that ␧

dƒ dx

冏

⫹

x⫽x*

冏

␧2 d 2ƒ

2! dx

2

⫹䡠䡠䡠⫹

x⫽x*

冏

␧n dnƒ

n! dxn

⫹ On⫹1(␧) ⭓ 0

(2.3)

x⫽x*

For ␧ sufficiently small, the first term will dominate the others, and since ␧ can be chosen both positive and negative, it follows that inequality (2.3) will hold only when

冏

dƒ dx

⫽0

(2.4)

x⫽x*

Continuing this argument further, inequality (2.3) will be true only when

38


冏

d 2ƒ dx 2

⭓0

(2.5)

x⫽x*

The same construction applies in the case of a local maximum but with inequality (2.2) reversed, and we obtain the following general result: Theorem 2.1 Necessary conditions for x* to be a local minimum (maximum) of ƒ on the open interval (a, b), providing that ƒ is twice differentiable, are that 1.

2.

冏冏

dƒ dx

d 2ƒ dx 2

⫽0

x⫽x*

⭓ 0 (⭐ 0)

x⫽x*

These are necessary conditions, which means that if they are not satisfied, then x* is not a local minimum (maximum). On the other hand, if they are satisfied, we still have no guarantee that x* is a local minimum (maximum). For example, consider the function ƒ(x) ⫽ x 3, shown in Figure 2.8. It satisfies the necessary conditions for both local minimum and local maximum at the origin, but the function does not achieve a minimum or a maximum at x* ⫽ 0. Definitions A stationary point is a point x* at which dƒ dx

冏

⫽0

x⫽x*

An inflection point or saddlepoint is a stationary point that does not correspond to a local optimum (minimum or maximum). To distinguish whether a stationary point corresponds to a local minimum, a local maximum, or an inflection point, we need the sufficient conditions of optimality. Theorem 2.2 Suppose at a point x* the first derivative is zero and the first nonzero higher order derivative is denoted by n.

2.2

OPTIMALITY CRITERIA

39

(i) If n is odd, then x* is a point of inflection. (ii) If n is even, then x* is a local optimum. Moreover: (a) If that derivative is positive, then the point x* is a local minimum. (b) If that derivative is negative, then the point x* is a local maximum. Proof This result is easily verified by recourse to the Taylor series expansion given in Eq. (2.1). Since the first nonvanishing higher order derivative is n, Eq. (2.1) reduces to ƒ(x* ⫹ ␧) ⫺ ƒ(x*) ⫽

冏

␧n dnƒ

n! dxn

⫹ On⫹1(␧)

(2.6)

x⫽x*

If n is odd, then the right-hand side of Eq. (2.6) can be made positive or negative by choosing ␧ positive or negative. This implies that depending on the sign of ␧, ƒ(x* ⫹ ␧) ⫺ ƒ(x*) could be positive or negative. Hence the function does not attain a minimum or a maximum at x*, and x* is an inflection point. Now, consider the case when n is even. Then the term ␧n is always positive, and for all ␧ sufficiently small the sign of Eq. (2.6) will be dominated by the first term. Hence, if (dnƒ/dxn)兩x⫽x* is positive, ƒ(x* ⫹ ␧) ⫺ ƒ(x*) ⬎ 0, and x* corresponds to a local minimum. A similar argument can be applied in the case of a local maximum. Applying Theorem 2.2 to the function ƒ(x) ⫽ x 3 shown in Figure 2.8, we note that dƒ dx

冏

x⫽0

⫽0

d 2ƒ dx 2

冏

x⫽0

⫽0

d 3ƒ dx 3

冏

⫽6

x⫽0

Figure 2.8. Illustration of inflection point.

40


Thus the first nonvanishing derivative is 3 (odd), and x ⫽ 0 is an inflection point. Remark In the above discussion, we have assumed that the function is always differentiable, in other words, that continuous first derivatives always exist. However, if the function is not differentiable at all points, then even the necessary condition for an unconstrained optimum, that the point is a stationary point, may not hold. For example, consider the piecewise linear function given by ƒ(x) ⫽

再

x 4⫺x

for x ⭐ 2 for x ⭓ 2

The function is continuous at all points. But it is not differentiable at the point x ⫽ 2, and the function attains its maximum at x ⫽ 2, which is not a stationary point by our definition. Example 2.1

Consider the function 165 4 3 ƒ(x) ⫽ 5x 6 ⫺ 36x 5 ⫹ ––– 2 x ⫺ 60x ⫹ 36

defined on the real line. The first derivative of this function is dƒ ⫽ 30x 5 ⫺ 180x 4 ⫹ 330x 3 ⫺ 180x 2 ⫽ 30x 2(x ⫺ 1)(x ⫺ 2)(x ⫺ 3) dx Clearly, the first derivative vanishes at x ⫽ 0, 1, 2, 3, and hence these points can be classified as stationary points. The second derivative of ƒ is d 2ƒ ⫽ 150x 4 ⫺ 720x 3 ⫹ 990x 2 ⫺ 360x dx 2 Evaluating this derivative at the four candidate points x ⫽ 0, 1, 2, 3, we obtain x

ƒ(x)

d 2ƒ/dx 2

0 1 2 3

36 27.5 44 5.5

0 60 ⫺120 540

2.2

OPTIMALITY CRITERIA

41

We conclude that x ⫽ 1, 3 are local minima and x ⫽ 2 is a local maximum. To characterize x ⫽ 0, the next higher derivative must be evaluated: d 3ƒ ⫽ 600x 3 ⫺ 2160x 2 ⫹ 1980x ⫺ 360 ⫽ ⫺360 dx 3 Since this is an odd-order derivative and is nonzero, the point x ⫽ 0 is not an optimum point but an inflection point. The next question is how to determine the global maximum or minimum of a function of one variable. Since the global optimum has to be a local optimum, a simple approach is to compute all local optima and choose the best. An algorithm based on this is given below: Maximize

ƒ(x)

Subject to a ⭐ x ⭐ b where a and b are practical limits on the values of the variable x. Once a function is bounded in an interval, you should notice that in addition to the stationary points, the boundary points can also qualify for the local optimum. Step 1. Set dƒ/dx ⫽ 0 and compute all stationary points. Step 2. Select all stationary points that belong to the interval [a, b]. Call them x1, x2, . . . , xN. These points, along with a and b, are the only points that can qualify for a local optimum. Step 3. Find the largest value of ƒ(x) out of ƒ(a), ƒ(b), ƒ(x1), . . . , ƒ(xN). This value becomes the global maximum point. Note: We did not try to classify the stationary points as local minimum, local maximum, or inflection points, which would involve the calculation of higher order derivatives. Instead it is easier to just compute their functional values and discard them. Example 2.2

Maximize ƒ(x) ⫽ ⫺x 3 ⫹ 3x 2 ⫹ 9x ⫹ 10 in the interval ⫺2 ⭐ x ⭐ 4. Set dƒ ⫽ ⫺3x 2 ⫹ 6x ⫹ 9 ⫽ 0 dx Solving this equation, we get x ⫽ 3 and x ⫽ ⫺1 as the two stationary points, and both are in the search interval.

42


To find the global maximum, evaluate ƒ(x) at x ⫽ 3, ⫺1, ⫺2, and 4: ƒ(3) ⫽ 37 ƒ(⫺2) ⫽ 12

ƒ(⫺1) ⫽ 5 ƒ(4) ⫽ 30

Hence x ⫽ 3 maximizes ƒ over the interval (⫺2, 4). Instead of evaluating all the stationary points and their functional values, we could use certain properties of the function to determine the global optimum faster. At the end of Section 2.1, we introduced unimodal functions, for which a local optimum is the global optimum. Unfortunately, the definition of unimodal functions does not suggest a simple test for determining whether or not a function is unimodal. However, there exists an important class of unimodal functions in optimization theory, known as convex and concave functions, which can be identified with some simple tests. A review of convex and concave functions and their properties is given in Appendix B. Example 2.3

Let us examine the properties of the function ƒ(x) ⫽ (2x ⫹ 1)2(x ⫺ 4) ƒ⬘(x) ⫽ (2x ⫹ 1)(6x ⫺ 15) ƒⴖ(x) ⫽ 24(x ⫺ 1) For all x ⭐ 1, ƒⴖ(x) ⭐ 0, and the function is concave in this region. For all x ⭓ 1, ƒⴖ(x) ⭓ 0, and the function is convex in the region. Note that the function has two stationary points, x ⫽ ⫺–12 and x ⫽ –52 . ƒⴖ(⫺–12 ) ⬍ 0, and the function has a local maximum at x ⫽ ⫺–12 . At x ⫽ –52 , ƒⴖ(–52 ) ⬎ 0, and the function attains a local minimum at this point. If we restricted the region of interest to x ⭐ 1, then ƒ(x) attains the global maximum at x ⫽ ⫺–12 since ƒ(x) is concave in this region and x ⫽ ⫺–12 is a local maximum. Similarly, if we restricted the region of interest to x ⭓ 1, then ƒ(x) attains the global minimum at x ⫽ –52 . However, over the entire range of x from ⫺⬁ to ⫹⬁, ƒ(x) has no finite global maximum or minimum. Example 2.4

Inventory Control

Many firms maintain an inventory of goods to meet future demand. Among the reasons for holding inventory is to avoid the time and cost of constant replenishment. On the other hand, to replenish only infrequently would imply large inventories that would tie up unnecessary capital and incure huge storage

2.2

OPTIMALITY CRITERIA

43

costs. Determining the optimal size of the inventory is a classic optimization problem, and a frequently used model in inventory control is the economic order quantity (EOQ) model. The EOQ model assumes that there is a steady demand of ␭ units per year for the product. Frequent replenishment is penalized by assuming that there is a setup or ordering cost of $K each time an order is placed, irrespective of how many units are ordered. The acquisition cost of each unit is $c. Excessive inventory is penalized by assuming that each unit will cost $h to store for one year. To keep things simple, we will assume that all demand must be met immediately (i.e., no back orders are allowed) and that replenishment occurs instantaneously as soon as an order is sent. Figure 2.9 graphically illustrates the change in the inventory level with respect to time. Starting at any time A with an inventory of B, the inventory level will decrease at the rate of ␭ units per unit time until it becomes zero at time C, when a fresh order is placed. The triangle ABC represents one inventory cycle that will be repeated throughout the year. The problem is to determine the optimal order quantity B, denoted by variable Q, and the common length of time C ⫺ A, denoted by T, between reorders. Since T is just the length of time required to deplete Q units at rate ␭, we get T⫽

Q ␭

The only remaining problem is to determine Q. Note that when Q is small, T will be small, implying more frequent reorders during the year. This will result in higher reorder costs but lower inventory holding cost. On the other hand, a large inventory (large Q) will result in a higher inventory cost but lower reorder costs. The basic inventory control problem is to determine the optimal value of Q that will minimize the sum of the inventory cost and reorder cost in a year. We shall now develop the necessary mathematical expression to optimize the yearly cost (cost/cycle ⫻ number of cycles/year).

Figure 2.9. Inventory problem.

44


Number of cycles (reorders)/year ⫽

1 ␭ ⫽ T Q

Cost per cycle ⫽ reorder cost ⫹ inventory cost

冉冊冉冊 Q hT 2

⫽ (K ⫹ cQ) ⫹ ⫽ K ⫹ cQ ⫹

hQ2 2␭

Note: The inventory cost per cycle is simply the cost of holding an average inventory of Q/2 for a length of time T. Thus, the yearly cost to be minimized is ƒ(Q) ⫽ ƒ⬘(Q) ⫽ ƒⴖ(Q) ⫽

␭K hQ ⫹ ␭c ⫹ Q 2 ⫺␭K 2

Q

⫹

h 2

2␭K ⬎0 Q3

for all Q ⬎ 0

Hence ƒ(Q) is a convex function, and if there exists a positive Q* such that ƒ⬘(Q*) ⫽ 0, then Q* minimizes ƒ(Q). Solving ƒ⬘(Q) ⫽ 0, we get Q* ⫽

冪2␭hK ⬎ 0

Thus, the optimal order quantity is Q* ⫽

冪2␭hK

and T* ⫽ time between reorders ⫽

冪2Kh␭

The parameter Q* is the famous economic order quantity, or EOQ, used frequently in inventory control.

2.3

2.3

REGION ELIMINATION METHODS

45


In Section 2.2, we addressed the static question, namely, how to determine whether or not a given solution is optimal. For this we developed a set of necessary and sufficient optimality conditions that a solution must satisfy to be optimal. Now we shall address the dynamic question, namely, how to determine the optimal or candidate optimal solutions. For this we shall develop a number of single-variable search methods for locating the optimal point in a given interval. Search methods that locate a single-variable optimum by successively eliminating subintervals so as to reduce the remaining interval of search are called region elimination methods. In Section 2.1, we introduced the definition of unimodal functions. unimodality is an extremely important functional property and virtually all the single-variable search methods currently in use require at least the assumption that within the domain of interest the function is unimodal. The usefulness of this property lies in the fact that if ƒ(x) is unimodal, then it is only necessary to compare ƒ(x) at two different points to predict in which of the subintervals defined by those two points the optimum does not lie. Theorem 2.3 Suppose ƒ is strictly unimodal† on the interval a ⭐ x ⭐ b with a minimum at x*. Let x1 and x2 be two points in the interval such that a ⬍ x1 ⬍ x2 ⬍ b. Comparing the functional values at x1 and x2, we can conclude: (i) If ƒ(x1) ⬎ ƒ(x2), then the minimum of ƒ(x) does not lie in the interval (a, x1). In other words, x* 僆 (x1, b) (see Figure 2.10). (ii) If ƒ(x1) ⬍ ƒ(x2), then the minimum does not lie in the interval (x2, b) or x* 僆 (a, x2) (see Figure 2.10).

Figure 2.10. Case (i) and case (ii) of Theorem 2.3.

† A function is strictly unimodal if it is unimodal and has no intervals of finite length in which the function is of constant value.

46


Proof Consider case (i), where ƒ(x1) ⬎ ƒ(x2). Suppose we assume the contrary, that a ⭐ x* ⭐ x1. Since x* is the minimum point, by definition we have ƒ(x*) ⭐ ƒ(x) for all x 僆 (a, b). This implies that ƒ(x*) ⭐ ƒ(x1) ⬎ ƒ(x2)

with x* ⬍ x1 ⬍ x2

But this is impossible, because the function has to be monotonic on either side of x* by the unimodality of ƒ(x). Thus, the theorem is proved by contradiction. A similar argument holds for case (ii). Note: When ƒ(x1) ⫽ ƒ(x2), we could eliminate both ends, (a, x1) and (x2, b), and the minimum must occur in the interval (x1, x2) providing ƒ(x) is strictly unimodal. By means of Theorem 2.3, sometimes called the elimination property, one can organize a search in which the optimum is found by recursively eliminating sections of the initial bounded interval. When the remaining subinterval is reduced to a sufficiently small length, the search is terminated. Note that without the elimination property, nothing less than an exhaustive search would suffice. The greatest advantage of these search methods is that they require only functional evaluations. The optimization functions need not be differentiable. As a matter of fact, the functions need not be in a mathematical or analytical form. All that is required is that given a point x the value of the function ƒ(x) can be determined by direct evaluation or by a simulation experiment. Generally, these search methods can be broken down into two phases: Bounding Phase. An initial coarse search that will bound or bracket the optimum. Interval Refinement Phase. A finite sequence of interval reductions or refinements to reduce the initial search interval to desired accuracy. 2.3.1

Bounding Phase

In the initial phase, starting at some selected trial point, the optimum is roughly bracketed within a finite interval by using the elimination property. Typically, this bounding search is conducted using some heuristic expanding pattern, although extrapolation methods have also been devised. An example of an expanding pattern is Swann’s method [1], in which the (k ⫹ 1)st test point is generated using the recursion xk⫹1 ⫽ xk ⫹ 2k⌬

for k ⫽ 0, 1, 2, . . .

2.3


47

where x0 is an arbitrarily selected starting point and ⌬ is a step-size parameter of suitably chosen magnitude. The sign of ⌬ is determined by comparing ƒ(x0), ƒ(x0 ⫹ 兩⌬兩), and ƒ(x0 ⫺ 兩⌬兩). If ƒ(x0 ⫺ 兩⌬兩) ⭓ ƒ(x0) ⭓ ƒ(x0 ⫹ 兩⌬兩) then, because of the unimodality assumption, the minimum must lie to the right of x0, and ⌬ is chosen to be positive. If the inequalities are reversed, ⌬ is chosen to be negative; if ƒ(x0 ⫺ 兩⌬兩) ⭓ ƒ(x0) ⭐ ƒ(x0 ⫹ 兩⌬兩) the minimum has been bracketed between x0 ⫺ 兩⌬兩 and x0 ⫹ 兩⌬兩 and the bounding search can be terminated. The remaining case, ƒ(x0 ⫺ 兩⌬兩) ⭐ ƒ(x0) ⭓ ƒ(x0 ⫹ 兩⌬兩) is ruled out by the unimodality assumption. However, occurrence of the above condition indicates that the given function is not unimodal. Example 2.5

Consider the problem of minimizing ƒ(x) ⫽ (100 ⫺ x)2 given the starting point x0 ⫽ 30 and a step size 兩⌬兩 ⫽ 5. The sign of ⌬ is determined by comparing ƒ(x0) ⫽ ƒ(30) ⫽ 4900 ƒ(x0 ⫹ 兩⌬兩) ⫽ ƒ(35) ⫽ 4225 ƒ(x0 ⫺ 兩⌬兩) ⫽ ƒ(25) ⫽ 5625 Since ƒ(x0 ⫺ 兩⌬兩) ⭓ ƒ(x0) ⭓ ƒ(x0 ⫹ 兩⌬兩) ⌬ must be positive, and the minimum point x* must be greater than 30. Thus, x1 ⫽ x0 ⫹ ⌬ ⫽ 35. Next,

x2 ⫽ x1 ⫹ 2⌬ ⫽ 45 ƒ(45) ⫽ 3025 ⬍ ƒ(x1)

48


therefore, x* ⬎ 35; x3 ⫽ x2 ⫹ 22⌬ ⫽ 65 ƒ(65) ⫽ 1225 ⬍ ƒ(x2) therefore, x* ⬎ 45; x4 ⫽ x3 ⫹ 23⌬ ⫽ 105 ƒ(105) ⫽ 25 ⬍ ƒ(x3) therefore, x* ⬎ 65; x5 ⫽ x4 ⫹ 24⌬ ⫽ 185 ƒ(185) ⫽ 7225 ⬎ ƒ(x4) therefore, x* ⬍ 185. Consequently, in six evaluations x* has been bracketed within the interval 65 ⭐ x* ⭐ 185 Note that the effectiveness of the bounding search depends directly on the step size ⌬. If ⌬ is large, a poor bracket, that is, a large initial interval, is obtained. On the other hand, if ⌬ is small, many evaluations may be necessary before a bound can be established. 2.3.2

Interval Refinement Phase

Once a bracket has been established around the optimum, then more sophisticated interval reduction schemes can be applied to obtain a refined estimate of the optimum point. The amount of subinterval eliminated at each step depends on the location of the trial points x1 and x2 within the search interval. Since we have no prior knowledge of the location of the optimum, it is reasonable to expect that the location of the trial points ought to be such that regardless of the outcome the interval should be reduced the same amount. Moreover, in the interest of efficiency, that same amount should be as large as possible. This is sometimes called the minimax criterion of search strategy. Interval Halving. This method deletes exactly one-half the interval at each stage. This is also called a three-point equal-interval search since it works with three equally spaced trial points in the search interval. The basic steps

2.3


49

of the search procedure for finding the minimum of a function ƒ(x) over the interval (a, b) are as follows: Step 1. Let xm ⫽ –12 (a ⫹ b) and L ⫽ b ⫺ a. Compute ƒ(xm). Step 2. Set x1 ⫽ a ⫹ –14 L and x2 ⫽ b ⫺ –14 L. Note that the points x1, xm, and x2 are all equally spaced at one-fourth the interval. Compute ƒ(x1) and ƒ(x2). Step 3. Compare ƒ(x1) and ƒ(xm). (i) If ƒ(x1) ⬍ ƒ(xm), then drop the interval (xm, b) by setting b ⫽ xm. The midpoint of the new search interval will now be x1. Hence, set xm ⫽ x1. Go to step 5. (ii) If ƒ(x1) ⭓ ƒ(xm), go to step 4. Step 4. Compare ƒ(x2) and ƒ(xm). (i) If ƒ(x2) ⬍ ƒ(xm), drop the interval (a, xm) by setting a ⫽ xm. Since the midpoint of the new interval will now be x2, set xm ⫽ x2. Go to step 5. (ii) If ƒ(x2) ⭓ ƒ(xm), drop the interval (a, x1) and (x2, b). Set a ⫽ x1 and b ⫽ x2. Note that xm continues to be the midpoint of the new interval. Go to step 5. Step 5. Compute L ⫽ b ⫺ a. If 兩L兩 is small, terminate. Otherwise return to step 2. Remarks 1. At each stage of the algorithm, exactly half the length of the search interval is deleted. 2. The midpoint of subsequent intervals is always equal to one of the previous trial points x1, x2, or xm. Hence, at most two functional evaluations are necessary at each subsequent step. 3. After n functional evaluations, the initial search interval will be reduced to (–12 )n / 2. 4. It has been shown by Kiefer [2] that out of all equal-interval searches (two-point, three-point, four-point, etc.), the three-point search or interval halving is the most efficient. Example 2.6

Interval Halving

Minimize ƒ(x) ⫽ (100 ⫺ x)2 over the interval 60 ⭐ x ⭐ 150. Here a ⫽ 60, b ⫽ 150, and L ⫽ 150 ⫺ 60 ⫽ 90. xm ⫽ –21 (60 ⫹ 150) ⫽ 105

50


Stage 1 90 x1 ⫽ a ⫹ –14 L ⫽ 60 ⫹ –– 4 ⫽ 82.5 90 x2 ⫽ b ⫺ –14 L ⫽ 150 ⫺ –– 4 ⫽ 127.5

ƒ(82.5) ⫽ 306.25 ⬎ ƒ(105) ⫽ 25 ƒ(127.5) ⫽ 756.25 ⬎ ƒ(105) Hence, drop the intervals (60, 82.5) and (127.5, 150). The length of the search interval is reduced from 90 to 45. Stage 2 a ⫽ 82.5

b ⫽ 127.5

xm ⫽ 105

L ⫽ 127.5 ⫺ 82.5 ⫽ 45 45 x1 ⫽ 82.5 ⫹ –– 4 ⫽ 93.75 45 x2 ⫽ 127.5 ⫺ –– 4 ⫽ 116.25

ƒ(93.75) ⫽ 39.06 ⬎ ƒ(105) ⫽ 25 ƒ(116.25) ⫽ 264.06 ⬎ ƒ(105) Hence, the interval of uncertainty is (93.75, 116.25). Stage 3 a ⫽ 93.75

b ⫽ 116.25

xm ⫽ 105

L ⫽ 116.25 ⫺ 93.75 ⫽ 22.5 x1 ⫽ 99.375 x2 ⫽ 110.625 ƒ(x1) ⫽ 0.39 ⬍ ƒ(105) ⫽ 25 Hence, delete the interval (105, 116.25). The new interval of uncertainty is now (93.75, 105), and its midpoint is 99.375 (old x1). Thus, in three stages (six functional evaluations), the initial search interval of length 90 has been reduced exactly to (90)(–12 )3 ⫽ 11.25.

2.3


51

Golden Section Search. In our discussion of region elimination methods and

minimax search strategies, we have observed the following: 1. If only two trials are available, then it is best to locate them equidistant from the center of the interval. 2. In general, the minimax strategy suggests that trials be placed symmetrically in the interval so that the subinterval eliminated will be of the same length regardless of the outcome of the trials. 3. The search scheme should require the evaluation of only one new point at each step. With this goal in mind, consider the symmetric placement of two trial points shown in Figure 2.11. Starting with a unit interval (purely for convenience), two trials are located a fraction ␶ from either end. With this symmetric placement, regardless of which of the corresponding function values is smaller, the length of the remaining interval is always ␶. Suppose that the right-hand subinterval is eliminated. It is apparent from Figure 2.12 that the remaining subinterval of length ␶ has the one old trial located interior to it at a distance 1 ⫺ ␶ from the left endpoint. To retain the symmetry of the search pattern, the distance 1 ⫺ ␶ should correspond to a fraction ␶ of the length of interval (which itself is of length ␶). With this choice of ␶, the next evaluation can be located at a fraction ␶ of the length of the interval from the right-hand endpoint (Figure 2.13). Hence, with the choice of ␶ satisfying 1 ⫺ ␶ ⫽ ␶ 2, the symmetric search pattern of Figure 2.11 is retained in the reduced interval of Figure 2.13. The solution of the quadratic equation is ␶⫽

⫺1 Ⳳ 兹5

2

the positive component of which is ␶ ⫽ 0.61803 . . . . The search scheme for locating trial points based upon this ratio is known as the golden section search. Note that after the first two evaluations, each succeeding evaluation will eliminate 1 ⫺ ␶ of the remaining interval. Hence, the interval remaining after N evaluations, assuming the initial interval was of unit length, will be

Figure 2.11. Golden section search.

52


Figure 2.12. Golden section intervals.

of length ␶ N⫺1. It can be shown that this is the asymptotic form of the optimum minimax search. Example 2.7

Golden Section Search

Consider again the problem of Example 2.6: Minimize ƒ(x) ⫽ (100 ⫺ x)2 over the bounded interval 60 ⭐ x ⭐ 150 For convenience, we rescale the interval to unit length by defining w⫽

x ⫺ 60 90

Thus, our problem becomes Minimize

ƒ(w) ⫽ (40 ⫺ 90w)2

Subject to

0⭐w⭐1

Step 1. I1 ⫽ (0, 1); L1 ⫽ 1. The first two evaluations are at w1 ⫽ ␶ ⫽ 0.618

with ƒ(w1) ⫽ 244.0

w2 ⫽ 1 ⫺ ␶ ⫽ ␶ 2 ⫽ 0.382

with ƒ(w2) ⫽ 31.6

Since ƒ(w2) ⬍ ƒ(w1) and w2 ⬍ w1, the interval w ⭓ w1 is eliminated. Step 2. I2 ⫽ (0, 0.618); L2 ⫽ 0.618 ⫽ ␶. The next evaluation is located at w3 ⫽ ␶ ⫺ ␶ 2 ⫽ ␶ (1 ⫺ ␶) ⫽ ␶ 3 ⫽ 0.236

with ƒ(w3) ⫽ 352

Since ƒ(w3) ⬎ ƒ(w2) and w3 ⬍ w2, the interval w ⭐ w3 is eliminated.

Figure 2.13. Golden section symmetry.

2.3


53

Step 3. I3 ⫽ (0.236, 0.618), L3 ⫽ 0.382 ⫽ ␶ 2. The next evaluation is at a distance ␶ times the length of the remaining interval from the current left endpoint or, equivalently, a distance 1 ⫺ ␶ times the interval length from the current right endpoint. Thus, w4 ⫽ 0.618 ⫺ (1 ⫺ ␶)L3 ⫽ 0.618 ⫺ ␶ 2L3 ⫽ 0.618 ⫺ ␶ 2(␶ 2) ⫽ 0.618 ⫺ ␶ 4 ⫽ 0.472

with ƒ(w4) ⫽ 6.15 Since ƒ(w4) ⬍ ƒ(w2) and w4 ⬎ w2, the interval w ⭐ w2 is eliminated. At this point the remaining interval of uncertainty is 0.382 ⭐ w ⭐ 0.618 in the w variable, or 94.4 ⭐ x ⭐ 115.6 in the original variable. If the search is continued through six evaluations, then the final interval in w will be ␶ N⫺1 ⫽ ␶ 5 ⫽ 0.09

which corresponds to an interval of 8.1 in variable x. This is a substantial improvement over the 11.25 achieved with interval halving. In general, if the left and right endpoints of the remaining interval of uncertainty (indicated by XR and XL) are saved, then all subsequent golden section trials can be calculated as either w ⫽ XR ⫺ ␶ n

or

w ⫽ XL ⫹ ␶ n

depending upon whether the left or right subinterval was last eliminated. In the above expressions, ␶ n indicates ␶ raised to the nth power, where n is the evaluation index. The golden section iterations can be terminated either by specifying a limit on the number of evaluations (and hence on the accuracy in the variable) or by specifying a relative accuracy in the function value. Preferably both tests should be applied. 2.3.3

Comparison of Region Elimination Methods

Let us now compare the relative efficiencies of the region elimination methods we have discussed so far. Let the original interval of uncertainty be denoted L1 and the final interval of uncertainty after N functional evaluations be LN.

54


Suppose we consider the fractional reduction (FR) of the original interval as a measure of merit of region elimination methods. Then FR(N) ⫽

LN L1

Recall that for the interval-halving and golden section search methods the lengths of the final intervals are given by L1(0.5)N / 2 and L1(0.618)N⫺1, respectively. Hence, the fractional reduction after N functional evaluations becomes FR(N) ⫽

再

(0.5)N / 2 (0.618)N⫺1

for interval halving for golden section

For the purpose of comparison, let us also include an exhaustive (brute-force) search, wherein the function is evaluated at N equally distant points [i.e., L1 is divided into N ⫹ 1 equal intervals of length L1 /(N ⫹ 1)]. Let x* be the point where the minimum of ƒ(x) is observed. Then the true minimum of ƒ(x) is bracketed in the interval

冋冉

x* ⫺

冊冉

冊册

L1 L1 , x* ⫹ N⫹1 N⫹1

In other words, LN ⫽ 2L1 /(N ⫹ 1). Hence, for an exhaustive search, FR(N) ⫽

2 N⫹1

Table 2.1 illustrates the values of FR(N) for the three search methods for selected values of N. From the table it is clear that the golden section search achieves the most fractional reduction of the original interval for the same number of functional evaluations. Looking at it another way, we can also compare the number of functional evaluations required for the same fractional

Table 2.1

Fractional Reduction Achieved

Number of Functional Evaluations Search Method

N⫽2

N⫽5

N ⫽ 10

N ⫽ 15

N ⫽ 20

Interval halving Golden section Exhaustive

0.5 0.618 0.667

0.177 0.146 0.333

0.031 0.013 0.182

0.006 0.001 0.125

0.0009 0.0001 0.095

2.4

POLYNOMIAL APPROXIMATION OR POINT ESTIMATION METHODS

55

reduction or accuracy. For a given FR(N) ⫽ E, the value of N may be computed as follows: For interval halving N⫽

2 ln(E) ln(0.5)

For golden section ln(E) ln(0.618)

N⫽1⫹ For exhaustive search N⫽

2 ⫺1 E

Table 2.2 gives the number of functional evaluations required for a desired accuracy in the determination of the minimum point. Once again, it is clear that the golden section search outperforms the other two methods by requiring the least number of functional evaluations for the same desired accuracy.

2.4 POLYNOMIAL APPROXIMATION OR POINT ESTIMATION METHODS

The region elimination methods discussed in the previous section require only that the function be unimodal. Hence, they are applicable to both continuous and discontinuous functions as well as to discrete variables. The entire search logic in these methods is based upon a simple comparison of function values at two trial points. Moreover, this comparison only takes account of the ordering of the function values and in no way involves the magnitude of the difference between the function values. The search methods discussed in this

Table 2.2

Required Number of Functional Evaluations

Accuracy Desired Search Method

E ⫽ 0.1

E ⫽ 0.05

E ⫽ 0.01

E ⫽ 0.001

Interval halving Golden section Exhaustive

7 6 19

9 8 39

14 11 199

20 16 1999

56


section do take into account the relative magnitudes of the function values and as a consequence often perform better than the region elimination methods. However, this gain in efficiency is achieved at the price of having to require that the functions optimized be sufficiently smooth. The basic motivating idea of the techniques to be presented here is that if the function is smooth, then it can be approximated by a polynomial, and the approximating polynomial can then be used to predict the location of the optimum. For this strategy to be effective, it is only necessary for the function in question to be both unimodal and continuous. The Weierstrass approximation theorem [3] guarantees that if the function is continuous on the interval, then it can be approximated as closely as desired by polynomials of sufficiently high order. Consequently, if the function is unimodal and a reasonably good approximating polynomial is at hand, then the function optimum can be reasonably well predicted by the polynomial optimum. As suggested by the Weierstrass theorem, improvements in the estimated optimum obtained through approximating polynomials can be achieved in two ways: by using a higher order polynomial or by reducing the interval over which the function is to be approximated. Of these, the second alternative is generally to be preferred, because beyond the third order the polynomial-fitting algebra becomes more complicated, and as a result of the unimodality assumption, interval reduction is rather easily accomplished. 2.4.1

Quadratic Estimation Methods

The simplest polynomial interpolation is a quadratic approximation. It is based upon the observation that if a function takes on its minimum in the interior of an interval, then it must be at least quadratic. If it is linear, it will assume its optimal value at one or the other of the endpoints of the interval. Thus, a quadratic estimation scheme assumes that within the bounded interval the function can be approximated by a quadratic and this approximation will improve as the points used to construct the approximation approach the actual minimum. Given three consecutive points x1, x2, x3 and their corresponding function values ƒ1, ƒ2, ƒ3, we seek to determine three constants a0, a1, and a2 such that the quadratic function q(x) ⫽ a0 ⫹ a1(x ⫺ x1) ⫹ a2(x ⫺ x1)(x ⫺ x2) agrees with ƒ(x) at these three points. We proceed by evaluating q(x) at each of the three given points. First of all, since ƒ1 ⫽ ƒ(x1) ⫽ q(x1) ⫽ a0 we have

2.4


57

a0 ⫽ ƒ 1 Next, since ƒ2 ⫽ ƒ(x2) ⫽ q(x2) ⫽ ƒ1 ⫹ a1(x2 ⫺ x1) we have a1 ⫽

ƒ2 ⫺ ƒ1 x2 ⫺ x1

Finally, at x ⫽ x3, ƒ3 ⫽ ƒ(x3) ⫽ q(x3) ⫽ ƒ1 ⫹

ƒ 2 ⫺ ƒ1 (x ⫺ x1) ⫹ a2(x3 ⫺ x1)(x3 ⫺ x2) x2 ⫺ x1 3

Solving for a2, we obtain a2 ⫽

冉

1 ƒ3 ⫺ ƒ1 ƒ2 ⫺ ƒ1 ⫺ x3 ⫺ x2 x3 ⫺ x1 x2 ⫺ x1

冊

Hence, given three points and their function values, the quadratic estimate can be constructed by simply evaluating the expressions for a0, a1, a2 given above. Now following the proposed strategy, if the approximating quadratic is a good approximation to the function to be optimized over the interval x1 to x3, then it can be used to predict the location of the optimum. Recall that the stationary points of a single-variable function can be determined by setting its first derivative to zero and solving for the roots of the resulting equation. In the case of our quadratic approximating function, dq ⫽ a1 ⫹ a2(x ⫺ x2) ⫹ a2(x ⫺ x1) ⫽ 0 dx can be solved to yield the estimate x⫽

x2 ⫹ x1 a ⫺ 1 2 2a2

Since the function ƒ(x) is unimodal over the interval in question and since the approximating quadratic is also a unimodal function, it is reasonable to expect that x will be a good estimate of the desired exact optimum x*.

58


Example 2.8

Quadratic Search

Consider the estimation of the minimum of ƒ(x) ⫽ 2x 2 ⫹

16 x

on the interval 1 ⭐ x ⭐ 5. Let x1 ⫽ 1, x3 ⫽ 5, and choose as x2 the midpoint x2 ⫽ 3. Evaluating the function, we obtain ƒ1 ⫽ 18

ƒ2 ⫽ 23.33

ƒ3 ⫽ 53.2

To calculate the estimate x, the constants a1 and a2 of the approximating function must be evaluated. Thus, a1 ⫽

23.33 ⫺ 18 8 ⫽ 3⫺1 3

a2 ⫽

1 53.2 ⫺ 18 8 46 ⫺ ⫽ 5⫺3 5⫺1 3 15

冉

冊

Substituting into the expression for x, x⫽

3⫹1 8/3 ⫺ ⫽ 1.565 2 2(46/15)

The exact minimum is x* ⫽ 1.5874. 2.4.2

Successive Quadratic Estimation Method

Developed by Powell [4], this method uses successive quadratic estimation. The algorithm is outlined below. Let x1 be the initial point and ⌬x the selected step size. Step 1. Compute x2 ⫽ x1 ⫹ ⌬x. Step 2. Evaluate ƒ(x1) and ƒ(x2). Step 3. If ƒ(x1) ⬎ ƒ(x2), let x3 ⫽ x1 ⫹ 2⌬x. If ƒ(x1) ⭐ ƒ(x2), let x3 ⫽ x1 ⫺ ⌬x. Step 4. Evaluate ƒ(x3) and determine Fmin ⫽ min{ƒ1, ƒ2, ƒ3} Xmin ⫽ point xi corresponding to Fmin

2.4


59

Step 5. Use the points x1, x2, x3 to calculate x using the quadratic estimation formula. Step 6. Check for termination. (a) Is Fmin ⫺ ƒ(x) small enough? (b) Is Xmin ⫺ x small enough? If both are satisfied, terminate. Otherwise, go to step 7. Step 7. Save the currently best point (Xmin or x), the two points bracketing it, or the two closest to it. Relabel them and go to step 4. Note that with the first pass through step 5 a bound of the minimum may not yet have been established. Consequently, the x obtained may be an extrapolation beyond x3. To ensure against a very large extrapolation step, a test should be added after step 5 that will replace x with a point calculated using a preset step size if x is very far from x3. Example 2.9

Powell’s Method

Consider the problem of Example 2.8: Minimize ƒ(x) ⫽ 2x 2 ⫹

16 x

with initial point x1 ⫽ 1 and step size ⌬x ⫽ 1. For convergence parameters use

冏

冏

冏

Difference in x ⭐ 3 ⫻ 10⫺2 x

冏

Difference in F ⭐ 3 ⫻ 10⫺3 F

Iteration 1 x2 ⫽ x1 ⫹ ⌬x ⫽ 2 ƒ(x1) ⫽ 18 ƒ(x2) ⫽ 16 ƒ(x1) ⬎ ƒ(x2); therefore set x3 ⫽ 1 ⫹ 2 ⫽ 3. ƒ(x3) ⫽ 23.33 Fmin ⫽ 16 Xmin ⫽ x2 16 ⫺ 18 Step 5. a1 ⫽ ⫽ ⫺2 2⫺1 5.33 1 23.33 ⫺ 18 a2 ⫽ ⫺ a1 ⫽ ⫹ 2 ⫽ 4.665 3⫺2 3⫺1 2 1⫹2 ⫺2 1 x⫽ ⫺ ⫽ 1.5 ⫹ ⫽ 1.714 2 2(4.665) 4.665 ƒ(x) ⫽ 15.210

Step Step Step Step

1. 2. 3. 4.

冉

冊

60


Step 6. Test for termination:

冏

冏

16 ⫺ 15.210 ⫽ 0.0519 ⬎ 0.003 15.210

Therefore continue. Step 7. Save x, the currently best point, and x1 and x2, the two points that bound it. Relabel the points in order and go to iteration 2, starting with step 4. Iteration 2 Step 4. x1 ⫽ 1 ƒ1 ⫽ 18 x2 ⫽ 1.714 ƒ2 ⫽ 15.210 ⫽ Fmin and Xmin ⫽ x2 x3 ⫽ 2 ƒ3 ⫽ 16 15.210 ⫺ 18 Step 5. a1 ⫽ ⫽ ⫺3.908 1.714 ⫺ 1 1.908 1 16 ⫺ 18 a2 ⫽ ⫺ (⫺3.908) ⫽ ⫽ 6.671 2 ⫺ 1.714 2 ⫺ 1 0.286 2.714 ⫺3.908 x⫽ ⫺ ⫽ 1.357 ⫹ 0.293 ⫽ 1.650 2 2(6.671) ƒ(x) ⫽ 15.142 Step 6. Test for termination:

冉

冏

冊

冏

15.210 ⫺ 15.142 ⫽ 0.0045 ⬎ 0.003 15.142

not satisfied

Step 7. Save x, the currently best point, and x1 ⫽ 1 and x2 ⫽ 1.714, the two points that bracket it. Iteration 3 Step 4. x1 ⫽ 1 ƒ1 ⫽ 18 x2 ⫽ 1.65 ƒ2 ⫽ 15.142 ⫽ Fmin and Xmin ⫽ x2 x3 ⫽ 1.714 ƒ3 ⫽ 15.210 15.142 ⫺ 18 Step 5. a1 ⫽ ⫽ ⫺4.397 1.65 ⫺ 1 1 15.210 ⫺ 18 a2 ⫽ ⫺ (⫺4.397) ⫽ 7.647 1.714 ⫺ 1.650 1.714 ⫺ 1 2.65 ⫺4.397 x⫽ ⫺ ⫽ 1.6125 2 2(7.647) ƒ(x) ⫽ 15.123

冉

冊

2.5

METHODS REQUIRING DERIVATIVES

61

Step 6. Test for termination: 15.142 ⫺ 15.123 (i) ⫽ 0.0013 ⬍ .003 15.123 (ii)

2.5

冏冏

冏

冏

1.65 ⫺ 1.6125 ⫽ 0.023 ⬍ 0.03 1.6125 Therefore, terminate iterations.


The search methods discussed in the preceding sections required the assumptions of unimodality and, in some cases, continuity of the performance index being optimized. It seems reasonable that, in addition to continuity, if it is assumed that the function is differentiable, then further efficiencies in the search could be achieved. Recall from Section 2.2 that the necessary condition for a point z to be a local minimum is that the derivative at z vanish, that is, ƒ⬘(z) ⫽ dƒ/dx兩x⫽z ⫽ 0. When ƒ(x) is a function of third-degree or higher terms involving x, direct analytic solution of ƒ⬘(x) ⫽ 0 would be difficult. Hence, a search method that successively approximates the stationary point of ƒ is required. We first describe a classical search scheme for finding the root of a nonlinear equation that was originally developed by Newton and later refined by Raphson [5]. 2.5.1

Newton–Raphson Method

The Newton–Raphson scheme requires that the function ƒ be twice differentiable. It begins with a point x1 that is the initial estimate or approximation to the stationary point or root of the equation ƒ⬘(x) ⫽ 0. A linear approximation of the function ƒ⬘(x) at the point x1 is constructed, and the point at which the linear approximation vanishes is taken as the next approximation. Formally, given the point xk to be the current approximation of the stationary point, the linear approximation of the function ƒ⬘(x) at xk is given by ƒ˜ ⬘(x; xk) ⫽ ƒ⬘(xk) ⫹ ƒⴖ(xk)(x ⫺ xk)

(2.7)

Setting Eq. (2.7) to be zero, we get the next approximation point as xk⫹1 ⫽ xk ⫺

ƒ⬘(xk) ƒⴖ(xk)

Figure 2.14 illustrates the general steps of Newton’s method. Unfortunately, depending on the starting point and the nature of the function, it is quite

62


Figure 2.14. Newton–Raphson method (convergence).

possible for Newton’s method to diverge rather than converge to the true stationary point. Figure 2.15 illustrates this difficulty. If we start at a point to the right of x0, the successive approximations will be moving away from the stationary point z. Example 2.10

Newton–Raphson Method

Consider the problem Minimize ƒ(x) ⫽ 2x 2 ⫹

16 x

Suppose we use the Newton–Raphson method to determine a stationary point of ƒ(x) starting at the point x1 ⫽ 1: ƒ⬘(x) ⫽ 4x ⫺

16 x2

ƒⴖ(x) ⫽ 4 ⫹

32 x3

Figure 2.15. Newton–Raphson method (divergence).

2.5


63

Step 1. x1 ⫽ 1 ƒ⬘(x1) ⫽ ⫺12 ƒⴖ(x1) ⫽ 36 ⫺12 x2 ⫽ 1 ⫺ ⫽ 1.33 36 Step 2. x2 ⫽ 1.33 ƒ⬘(x2) ⫽ ⫺3.73 ƒⴖ(x2) ⫽ 17.6 ⫺3.73 x3 ⫽ 1.33 ⫺ ⫽ 1.54 17.6 We continue until 兩ƒ⬘(xk)兩 ⬍ ␧, where ␧ is a prespecified tolerance. 2.5.2

Bisection Method

If the function ƒ(x) is unimodal over a given search interval, then the optimal point will be the one where ƒ⬘(x) ⫽ 0. If both the function value and the derivative of the function are available, then an efficient region elimination search can be conducted using just a single point rather than a pair of points to identify a point at which ƒ⬘(x) ⫽ 0. For example, if, at a point z, ƒ⬘(z) ⬍ 0, then assuming that the function is unimodal, the minimum cannot lie to the left of z. In other words, the interval x ⭐ z can be eliminated. On the other hand, if ƒ⬘(z) ⬎ 0, then the minimum cannot lie to the right of z, and the interval x ⭓ z can be eliminated. Based on these observations, the bisection method (sometimes called the Bolzano search) can be constructed. Determine two points L and R such that ƒ⬘(L) ⬍ 0 and ƒ⬘(R) ⬎ 0. The stationary point is between the points L and R. We determine the derivative of the function at the midpoint, z⫽

L⫹R 2

If ƒ⬘(z) ⬎ 0, then the interval (z, R) can be eliminated from the search. On the other hand, if ƒ⬘(z) ⬍ 0, then the interval (L, z) can be eliminated. We shall now state the formal steps of the algorithm. Given a bounded interval a ⭐ x ⭐ b and a termination criterion ␧: Step 1. Set R ⫽ b, L ⫽ a; assume ƒ⬘(a) ⬍ 0 and ƒ⬘(b) ⬎ 0. Step 2. Calculate z ⫽ (R ⫹ L)/2, and evaluate ƒ⬘(z). Step 3. If 兩ƒ⬘(z)兩 ⭐ ␧, terminate. Otherwise, if ƒ⬘(z) ⬍ 0, set L ⫽ z and go to step 2. If ƒ⬘(z) ⬎ 0, set R ⫽ z and go to step 2. Note that the search logic of this region elimination method is based purely on the sign of the derivative and does not use its magnitude. A method that uses both is the secant method, to be discussed next.

64

2.5.3


Secant Method

The secant method combines Newton’s method with a region elimination scheme for finding a root of the equation ƒ⬘(x) ⫽ 0 in the interval (a, b) if it exists. Suppose we are interested in finding the stationary point of ƒ(x) and we have two points L and R in (a, b) such that their derivatives are opposite in sign. The secant algorithm then approximates the function ƒ⬘(x) as a ‘‘secant line’’ (a straight line between the two points) and determines the next point where the secant line of ƒ⬘(x) is zero (see Figure 2.16). Thus, the next approximation to the stationary point x* is given by z⫽R⫺

ƒ⬘(R) [ƒ⬘(R) ⫺ ƒ⬘(L)]/(R ⫺ L)

If 兩ƒ⬘(z)兩 ⭐ ␧, we terminate the algorithm. Otherwise, we select z and one of the points L or R such that their derivatives are opposite in sign and repeat the secant step. For example, in Figure 2.16, we would have selected z and R as the next two points. It is easy to see that, unlike the bisection search, the secant method uses both the magnitude and sign of the derivatives and hence can eliminate more than half the interval in some instances (see Figure 2.16). Example 2.11

Secant Method

Consider again the problem of Example 2.10: Minimize ƒ(x) ⫽ 2x 2 ⫹ ƒ⬘(x) ⫽

16 over the interval 1 ⭐ x ⭐ 5 x

dƒ(x) 16 ⫽ 4x ⫺ 2 dx x

Figure 2.16. Secant method.

2.5


65

Iteration 1 Step 1. R ⫽ 5

L⫽1 ƒ⬘(R) ⫽ 19.36 19.36 Step 2. z ⫽ 5 ⫺ ⫽ 2.53 (19.36 ⫹ 12)/4 Step 3. ƒ⬘(z) ⫽ 7.62 ⬎ 0; set R ⫽ 2.53.

ƒ⬘(L) ⫽ ⫺12

Iteration 2 7.62 ⫽ 1.94 (7.62 ⫹ 12)/1.53 Step 3. ƒ⬘(z) ⫽ 3.51 ⬎ 0; set R ⫽ 1.94.

Step 2. z ⫽ 2.53 ⫺

Continue until 兩ƒ⬘(z)兩 ⭐ ␧.

2.5.4

Cubic Search Method

This is a polynomial approximation method in which a given function ƒ to be minimized is approximated by a third-order polynomial. The basic logic is similar to the quadratic approximation scheme. However, in this instance, because both the function value and the derivative value are available at each point, the approximating polynomial can be constructed using fewer points. The cubic search starts with an arbitrary point x1 and finds another point x2 by a bounding search such that the derivatives ƒ⬘(x1) and ƒ⬘(x2) are of opposite sign. In other words, the stationary point x where ƒ⬘(x) ⫽ 0 is bracketed between x1 and x2. A cubic approximation function of the form ƒ(x) ⫽ a0 ⫹ a1(x ⫺ x1) ⫹ a2(x ⫺ x1)(x ⫺ x2) ⫹ a3(x ⫺ x1)2(x ⫺ x2)

(2.8)

is fitted such that Eq. (2.8) agrees with ƒ(x) at the two points x1 and x2. The first derivative of ƒ(x) is given by dƒ(x) ⫽ a1 ⫹ a2(x ⫺ x1) ⫹ a2(x ⫺ x2) ⫹ a3(x ⫺ x1)2 ⫹ 2a3(x ⫺ x1)(x ⫺ x2) dx (2.9) The coefficients a0, a1, a2, and a3 of Eq. (2.8) can now be determined using the values of ƒ(x1), ƒ(x2), ƒ⬘(x1), and ƒ⬘(x2) by solving the following linear equations:

66


ƒ1 ⫽ ƒ(x1) ⫽ a0 ƒ2 ⫽ ƒ(x2) ⫽ a0 ⫹ a1(x2 ⫺ x1) ƒ1⬘ ⫽ ƒ⬘(x1) ⫽ a1 ⫹ a2(x1 ⫺ x2) ƒ⬘2 ⫽ ƒ⬘(x2) ⫽ a1 ⫹ a2(x2 ⫺ x1) ⫹ a3(x2 ⫺ x1)2 Note that the above system can be solved very easily in a recursive manner. Then, as in the quadratic case discussed earlier, given these coefficients, an estimate of the stationary point of ƒ can be obtained from the approximating cubic of Eq. (2.8). In this case, when we set the derivative of ƒ(x) given by Eq. (2.9) to zero, we get a quadratic equation. By applying the formula for the root of the quadratic equation, a closed-form solution to the stationary point x of the approximating cubic is obtained as follows:

冦

x2 x ⫽ x2 ⫺ ␮(x2 ⫺ x1) x1

if ␮ ⬍ 0 if 0 ⭐ ␮ ⭐ 1 if ␮ ⬎ 1

(2.10)

where ␮⫽

z⫽ w⫽

ƒ⬘2 ⫹ w ⫺ z ƒ⬘2 ⫺ ƒ1⬘ ⫹ 2w

冉再

冊

3(ƒ1 ⫺ ƒ2) ⫹ ƒ⬘1 ⫹ ƒ2⬘ x2 ⫺ x1 (z 2 ⫺ ƒ⬘1ƒ2⬘)1 / 2 ⫺(z 2 ⫺ ƒ⬘1ƒ⬘2)1 / 2

if x1 ⬍ x2 if x1 ⬎ x2

The definition of w in Eq. (2.10) ensures that the proper root of the quadratic equation is selected, while forcing ␮ to lie between 0 and 1 ensures that the predicted point x lies between the bounds x1 and x2. Once again we select the next two points for the cubic approximation as x and one of the points x1 or x2 such that the derivatives of the two points are of opposite sign and repeat the cubic approximation. Given an initial point x0, positive step size ⌬, and termination parameters ␧1 and ␧2, the formal steps of the cubic search algorithm are as follows: Step 1. Compute ƒ⬘(x0). If ƒ⬘(x0) ⬍ 0, compute xK⫹1 ⫽ xK ⫹ 2K⌬ for K ⫽ 0, 1, . . . . If ƒ⬘(x0) ⬎ 0, compute xK⫹1 ⫽ xK ⫺ 2K⌬ for K ⫽ 0, 1, 2, . . . . Step 2. Evaluate ƒ⬘(x) for points xK⫹1 for K ⫽ 0, 1, 2, . . . until a point xM is reached at which ƒ⬘(xM⫺1)ƒ⬘(xM) ⭐ 0.

2.5


67

Then set x1 ⫽ xM⫺1, x2 ⫽ xM. Compute ƒ1, ƒ2, ƒ⬘1, and ƒ2⬘. Step 3. Calculate the stationary point x of the cubic approximating function using Eq. (2.10). Step 4. If ƒ(x) ⬍ ƒ(x1), go to step 5. Otherwise, set x ⫽ x ⫹ –12 (x ⫺ x1) until ƒ(x) ⭐ ƒ(x1) is achieved. Step 5. Termination check: If 兩ƒ⬘(x)兩 ⭐ ␧1 and 兩(x ⫺ x1)/ x兩 ⭐ ␧2, stop. Otherwise, set (i) x2 ⫽ x1 and x1 ⫽ x if ƒ⬘(x)ƒ⬘(x1) ⬍ 0 (ii) x1 ⫽ x if ƒ⬘(x)ƒ⬘(x2) ⬍ 0 In either case, continue with step 3. Note that steps 1 and 2 constitute a bounding search proceeding along an expanding pattern in which the sign of the derivative is used to determine when the optimum has been overshot. Step 3 represents the calculation of the optimum of the approximating cubic. Step 4 is merely a safety check to ensure that the generated estimate is in fact an improvement. Provided that direct values of the derivative are available, this search is unquestionably the most efficient search method currently available. However, if derivatives must be obtained by differences, then the quadratic-based search of Powell is preferred. Example 2.12

Cubic Search with Derivatives

Consider again, Minimize ƒ(x) ⫽ 2x 2 ⫹

16 x

with the initial point x0 ⫽ 1 and step size ⌬ ⫽ 1. For convergence parameters use ␧1 ⫽ 10⫺2

ƒ⬘(x) ⫽

␧2 ⫽ 3 ⫻ 10⫺2

dƒ 16 ⫽ 4x ⫺ 2 dx x

Iteration 1 Step 1. ƒ⬘(1) ⫽ ⫺12 ⬍ 0. Hence, x1 ⫽ 1 ⫹ 1 ⫽ 2. Step 2. ƒ⬘(2) ⫽ 4 Since ƒ⬘(1)ƒ⬘(2) ⫽ ⫺48 ⬍ 0, a stationary point has been bracketed between 1 and 2. Set x1 ⫽ 1, x2 ⫽ 2. Then, ƒ1 ⫽ 18, ƒ2 ⫽ 16, ƒ1⬘ ⫽ ⫺12, ƒ⬘2 ⫽ 4.

68


Step 3. z ⫽ –31 (18 ⫺ 16) ⫹ (⫺12) ⫹ 4 ⫽ ⫺2 w ⫽ ⫹[4 ⫺ (⫺12)(4)]1 / 2 ⫽ (52)1 / 2 ⫽ 7.211 4 ⫹ 7.211 ⫺ (⫺2) ␮⫽ ⫽ 0.4343 4 ⫺ (⫺12) ⫹ 2(7.211) x ⫽ 2 ⫺ 0.4343(2 ⫺ 1) ⫽ 1.5657 Step 4. ƒ(1.5657) ⫽ 15.1219 ⬍ ƒ(x1) ⫽ 18 Therefore, continue. Step 5. Termination check: ƒ⬘(1.5657) ⫽ ⫺0.2640

clearly not terminated

Since ƒ⬘(x)ƒ⬘(x2) ⫽ (⫺0.2640)(4) ⬍ 0 set x1 ⫽ x ⫽ 1.5657. Iteration 2 3 (15.1219 ⫺ 16) ⫹ (⫺0.2640) ⫹ 4 ⫽ ⫺2.3296 0.4343 w ⫽ ⫹[(2.3296)2 ⫺ (⫺0.2640)(4)]1 / 2 ⫽ 2.5462 4 ⫹ 2.5462 ⫺ (⫺2.3296) ␮⫽ ⫽ 0.9486 4 ⫺ (⫺0.2640) ⫹ 2(2.5462) x ⫽ 2 ⫺ 0.9486(2 ⫺ 1.5657) ⫽ 1.5880 Step 4. ƒ(1.5880) ⫽ 15.1191 ⬍ ƒ(x1) ⫽ 15.1219 Therefore, continue. Step 5. Termination check: Step 3. z ⫽

冏

ƒ⬘(1.5880) ⫽ 0.0072 ⬍ 10⫺2

冏

1.5880 ⫺ 1.5657 ⫽ 0.0140 ⬍ 3 ⫻ 10⫺2 1.5880

Therefore, terminate. Note that using the same two points, x1 ⫽ 1 and x2 ⫽ 2, the quadratic search of Example 2.9 returned an estimate x ⫽ 1.714, whereas the cubic search produced 1.5657. The exact minimum is 1.5874, indicating the clear superiority of the higher order polynomial fit.

2.6

2.6

COMPARISON OF METHODS

69

COMPARISON OF METHODS

From a theoretical point of view, it can be shown that point estimation methods such as Powell’s quadratic search or the cubic search with derivatives are intrinsically superior to region elimination methods such as golden section. In particular, it has been shown that the quadratic scheme exhibits superlinear asymptotic convergence to the true minimum; that is, the deviation of the (k ⫹ 1)st estimate from the true minimum is proportional to the deviation of the kth estimate raised to the ␣ power, where ␣ ⬎ 1 (␣ ⫽ 1.3 has been claimed [6]). By contrast, if in the case of the golden section search the kth estimate of the true minimum is taken to be the midpoint of the interval remaining after k evaluations, then the deviation of this estimate from the actual minimum decreases linearly from the kth to the (k ⫹ 1)st iterations. Assuming comparable regions of convergence for both methods, this indicates that the quadratic search method will converge more rapidly than any region elimination scheme. Of course, this comparison is based on assuming comparable regions of convergence and on well-behaved unimodal functions. The limited numerical comparisons available in the literature do not indicate any overwhelming advantage of either the derivative-based or the quadratic or region elimination search method over the others. If functional evaluations are lengthy and costly in terms of computer time, Box et al. [7] claim superiority of a search strategy based on a modification of Powell’s method. This claim seems to be supported by the limited computational experiments presented by Himmelblau [8], who compares this strategy with one based on the golden section search and shows that a modified Powell’s search requires fewer functional evaluations to attain a specified accuracy in the estimate of the minimizing point. Certainly, if very high accuracy is desired, the polynomial approximation methods are quite superior. On the other hand, for strongly skewed or possibly multimodal functions, Powell’s search method has been known to converge at a much slower rate than region elimination schemes. Hence, if reliability is of prime importance, then golden section is the ideal choice. Because of these considerations, it is recommended that Powell-type search methods generally be used along with a golden section search to which the program can default if it encounters difficulties in the course of iterations. Readers may also refer to Brent [9] for a good discussion of the relative efficiencies of different search methods. In a small study conducted as a class project for the optimization course at Purdue University, the following point estimation methods were compared on the function ƒ(x) ⫽ sink x for different values of k: Bisection method Powell’s method Cubic search

70


In addition, the best region elimination method, namely the golden section search, was also included for comparison. Three criteria—solution time, solution accuracy, and sensitivity to convergence parameter—were used to evaluate these methods. The first two criteria were examined by varying the value of the power k. Odd powers between 1 and 79 were chosen for k. Note that the minimum point is always x* ⫽ 4.71239 and ƒ(x*) ⫽ ⫺1.0 for all values of k. However, as k increases, the function becomes less smooth and has narrow valleys near the minimum. This would make the point estimation methods slow and less accurate. The sensitivity criterion was tested by varying the convergence parameter. As expected, results showed an increase in solution times as k increased. The bisection method showed the greatest increase in solution times with respect to k, due to rapid increase in the values of the gradient close to the minimum. Golden section search was unaffected by the increase in power k. Similarly, solution accuracy, measured in terms of percentage error to the true minimum, became worse as k increased for all three methods—bisection, Powell, and cubic. However, golden section search was unaffected by changes in the steepness of the function. Sensitivity to change in the convergence criterion was minimal for all four methods tested. 2.7

SUMMARY

In this chapter, we discussed necessary and sufficient optimality conditions for functions of a single variable. We showed that a necessary condition for the optimum is that the point be a stationary point, that is, that the first derivative be zero. Second-order and higher order derivatives were then used to test whether a point corresponds to a minimum, a maximum, or an inflection point. Next we addressed the question of identifying candidate optima. We developed a number of search methods called region elimination methods for locating an optimal point in a given interval. We showed that the golden section search is generally the preferred algorithm due to its computational efficiency and ease of implementation. The region elimination method required a simple comparison of function values at two trial points and hence took into account the ordering of the function values only. To involve the magnitude of the difference between the function values as well, we developed point estimation methods. These involved a quadratic or cubic approximation of the function to determine the optimum. We pointed out that, assuming comparable regions of convergence, point estimation methods converge more rapidly than region elimination methods as a class for well-behaved unimodal functions. However, for strongly skewed or multimodal functions, the golden section search is more reliable. We concluded with a recommendation that Powell-type successive quadratic estimation search be generally used along with a golden section

PROBLEMS

71

search to which the program can default if it encounters difficulties in the course of iterations. REFERENCES 1. Swann, W. H., ‘‘Report on the Development of a Direct Search Method of Optimization,’’ ICI Ltd., Central Instr. Res. Lab, Res. Note, 64 / 3, London, 1964. 2. Kiefer, J., ‘‘Optimum Sequential Search and Approximation Methods under Minimum Regularity Assumptions,’’ J. Soc. Ind. Appl. Math., 5(3), 105–125 (1957). 3. Bartle, R., Elements of Real Analysis, Wiley, New York, 1976. 4. Powell, M. J. D., ‘‘An Efficient Method for Finding the Minimum of a Function of Several Variables without Calculating Derivatives,’’ Computer J., 7, 155–162 (1964). 5. Raphson, J., History of Fluxion, London, 1725. 6. Kowalik, J., and M. R. Osborne, Methods for Unconstrained Optimization Problems, American Elsevier, New York, 1968. 7. Box, M. J., D. Davies, and W. H. Swann, Nonlinear Optimization Techniques, ICI Monograph, Oliver and Boyd, Edinburgh, 1969. 8. Himmelblau, D. M., Applied Nonlinear Programming, McGraw-Hill, New York, 1972. 9. Brent, R. P., Algorithms for Minimization without Derivatives, Prentice-Hall, Englewood Cliffs, NJ, 1973. 10. Phillips, D. T., A. Ravindran, and J. J. Solberg, Operations Research: Principles and Practice, Wiley, New York, 1976.

PROBLEMS

2.1. What is an inflection point and how do you identify it? 2.2. How do you test a function to be convex or concave? 2.3. What is the unimodal property and what is its significance in singlevariable optimization? 2.4. Suppose a point satisfies sufficiency conditions for a local minimum. How do you establish that it is a global minimum? 2.5. Cite a condition under which a search method based on polynomial interpolation may fail. 2.6. Are region elimination methods as a class more efficient than point estimation methods? Why or why not? 2.7. In terminating search methods, it is recommended that both the difference in variable values and the difference in the function values be

72


tested. Is it possible for one test alone to indicate convergence to a minimum while the point reached is really not a minimum? Illustrate graphically. 2.8. Given the following functions of one variable: x3 (a) ƒ(x) ⫽ x 5 ⫹ x 4 ⫺ ⫹2 3 (b) ƒ(x) ⫽ (2x ⫹ 1)2(x ⫺ 4) Determine, for each of the above functions, the following: (i) Region(s) where the function is increasing; decreasing (ii) Inflexion points, if any (iii) Region(s) where the function is concave; convex (iv) Local and global maxima, if any (v) Local and global minima, if any 2.9. A cross-channel ferry is constructed to transport a fixed number of tons (L) across each way per day (see Figure 2.17). If the cost of construction of the ferry without the engines varies as the load (l) and the cost of the engines varies as the product of the load and the cube of the speed (v), show that the total cost of construction is least when twice as much money is spent on the ferry as on the engine. (You may neglect the time of loading and unloading and assume the ferry runs continuously.) 2.10. A forest fire is burning down a narrow valley of width 2 miles at a velocity of 32 fpm (Figure 2.18). A fire can be contained by cutting a fire break through the forest across the width of the valley. A man can clear 2 ft of the fire break in a minute. It costs $20 to transport each man to the scene of fire and back, and each man is paid $6 per hour while there. The value of timber is $2000 per square mile. How many men should be sent to fight the fire so as to minimize the total costs? 2.11. Consider the unconstrained optimization of a single-variable function ƒ(x). Given the following information about the derivatives of orders 1, 2, 3, 4 at the point xi (i ⫽ 1, 2, . . . , 10), identify the nature of the test point xi (i.e., whether it is maximum, minimum, inflection point, nonoptimal, no conclusion possible, etc.):

Figure 2.17. Problem 2.9 schematic.

73

PROBLEMS

Figure 2.18. Problem 2.10 schematic.

xi

ƒ⬘(xi)

ƒⴖ(xi)

ƒ⵮(xi)

ƒⴖⴖ(xi)

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10

0 0 0

⫹

Anything

Anything Anything Anything Anything Anything

⫺

⫺ ⫺

0 0 0 0

0 0 0 0

⫹

⫹ ⫺

0

⫹

0

Anything Anything ⫺

⫺

0 0 0 Anything

Anything

⫹

⫺

0 ⫹

2.12. State whether each of the following functions is convex, concave, or neither. (a) ƒ(x) ⫽ ex (b) ƒ(x) ⫽ e⫺x 1 (c) ƒ(x) ⫽ 2 x (d) ƒ(x) ⫽ x ⫹ log x for x ⬎ 0 (e) ƒ(x) ⫽ 兩x兩 (f) ƒ(x) ⫽ x log x for x ⬎ 0 2k (g) ƒ(x) ⫽ x where k is an integer (h) ƒ(x) ⫽ x 2k⫹1 where k is an integer 2.13. Consider the function ƒ(x) ⫽ x 3 ⫺ 12x ⫹ 3

over the region ⫺4 ⭐ x ⭐ 4

Determine the local minima, local maxima, global minimum, and global maximum of ƒ over the given region. 2.14. Identify the regions over which the following function is convex and where it is concave: ƒ(x) ⫽ e⫺x

2

Determine its global maximum and global minimum.

74


2.15. Consider a single-period inventory model for perishable goods as follows: Demand is a random variable with density ƒ; i.e., P(demand ⭐ x) ⫽ 兰x0ƒ(x) dx. All stockouts are lost sales. C is the unit cost of each item. p is the loss due to each stockout (includes loss of revenue and goodwill). r is the selling price per unit. l is the salvage value of each unsold item at the end of the period. The problem is to determine the optimal order quantity Q that will maximize the expected net revenue for the season. The expected net revenue, denoted by ⌸(Q), is given by

冕

Q

⌸(q) ⫽ r␮ ⫹ l

0

⫺ (p ⫹ r)

(Q ⫺ x)ƒ(x) dx

冕

⬁

Q

(x ⫺ Q)ƒ(x) dx ⫺ CQ

where ␮ ⫽ expected demand ⫽ 兰⬁0 xƒ(x) dx. (a) Show that ⌸(Q) is a concave function in Q(⭓ 0). (b) Explain how you will get an optimal ordering policy using (a). (c) Compute the optimal order policy for the example given below: C ⫽ $2.50

r ⫽ $5.00

l⫽0

p ⫽ $2.50

1 for 100 ⭐ x ⭐ 500 ƒ(x) ⫽ 400 0 otherwise Hint: Use the Leibniz rule for differentiation under the integral sign.

冦

2.16. Suppose we conduct a single-variable search using (a) the golden section method and (b) the bisection method with derivatives estimated numerically by differences. Which is likely to be more efficient? Why? 2.17. Carry out a single-variable search to minimize the function

ƒ(x) ⫽ 3x 2 ⫹

12 ⫺5 x3

on the interval –21 ⭐ x ⭐ –25

PROBLEMS

75

using (a) golden section, (b) interval halving, (c) quadratic-based search, and (d) cubic-derivative-based search. Each search method is to use four functional evaluations only. Compare the final search intervals obtained by the above methods. 2.18. Determine the minimum of ƒ(x) ⫽ (10x 3 ⫹ 3x 2 ⫹ x ⫹ 5)2 starting at x ⫽ 2 and using a step size ⌬ ⫽ 0.5. (a) Using region elimination: expanding pattern bounding plus six steps of golden section. (b) Using quadratic point estimation: three iterations of Powell’s method. 2.19. Determine the real roots of the equation (within one decimal point accuracy) ƒ(x) ⫽ 3000 ⫺ 100x 2 ⫺ 4x 5 ⫺ 6x 6 ⫽ 0 using (a) the Newton–Raphson method, (b) the bisection method, and (c) the secant method. You may find it helpful to write a small computer program to conduct these search methods. 2.20. (a) Explain how problem 2.19 can be converted to an unconstrained nonlinear optimization problem in one variable. (b) Write a computer program for the golden section search and solve the optimization problem formulated in part (a). Compare the solution with those obtained in problem 2.19. 2.21. An experimenter has obtained the following equation to describe the trajectory of a space capsule (Phillips, Ravindran, and Solberg [10]): ƒ(x) ⫽ 4x 3 ⫹ 2x ⫺ 3x 2 ⫹ e x / 2 Determine a root of the above equation using any of the derivativebased methods. 2.22. Minimize the following functions using any univariate search method up to one-decimal-point accuracy (Phillips, Ravindran, and Solberg [10]). (a) Minimize ƒ(x) ⫽ 3x 4 ⫹ (x ⫺ 1)2 over the range [0, 4]. (b) Minimize ƒ(x) ⫽ (4x)(sin x) over the range [0, ⌸]. 2 (c) Minimize ƒ(x) ⫽ 2(x ⫺ 3)2 ⫹ e0.5x over the range (0, 100).

76


2.23. In a chemical plant, the costs of pipes, their fittings, and pumping are important investment costs. Consider the design of a pipeline L feet long that should carry fluid at the rate of Q gpm. The selection of economic pipe diameter D (in.) is based on minimizing the annual cost of pipe, pump, and pumping. Suppose the annual cost of a pipeline with a standard carbon steel pipe and a motor-driven centrifugal pump can be expressed as ƒ ⫽ 0.45L ⫹ 0.245LD1.5 ⫹ 325(hp)1 / 2 ⫹ 61.6(hp)0.925 ⫹ 102 where hp ⫽ 4.4 ⫻ 10⫺8

LQ3 LQ2.68 ⫺9 ⫹ 1.92 ⫻ 10 D5 D4.68

Formulate the appropriate single-variable optimization problem for designing a pipe of length 1000 ft with a flow rate of 20 gpm. The diameter of the pipe should be between 0.25 and 6 in. Solve using the golden section search. 2.24. Solve the following problem by any univariate search method: Minimize

ƒ(x) ⫽ –13 兹x 2 ⫹ 25 ⫹

Subject to

0⭐x⭐6

6⫺x 5

2.25. You are interested in finding the positive root of the following equation: ƒ(x) ⫽ 3000 ⫺ 100x ⫺ 50x 2 ⫺ 4x 3 Given the initial bound for the root as (0, 10), apply one step of the secant method to find the first approximation to the root. What will be the new bound for the root? 2.26. John is planning to retire in 30 years with a nest egg of $500,000. He can save $2500 each year toward retirement. What minimum annual return should he get from his savings to reach his retirement goal? Hint: Annual savings of $1 for N years will accumulate to [(1 ⫹ i)N ⫺ 1]/i, where i is the compounded annual rate of return. 2.27. The XYZ Company plans to invest $10 million in a new manufacturing process that will result in an estimated after-tax benefit of $2 million over the 8-year service life of the process. Compute the annual rate of return for the investment.

PROBLEMS

77

Hint: The present value of $1 saving in year N at an annual interest rate i is given by P⫽

1 (1 ⫹ i)N

2.28. (a) Using golden section search and 10 functional evaluations, what is the interval reduction we can achieve? (b) To achieve the same interval reduction, how many function evaluations would be needed in the interval-halving method?

3 FUNCTIONS OF SEVERAL VARIABLES In this chapter we examine fundamental concepts and useful methods for finding minima of unconstrained functions of several variables. We build and expand upon the material given in Chapter 2, recognizing the value of the methods considered there in the discussion to follow. With the aid of a few necessary mathematical notational conventions and results from linear analysis and the calculus (see Appendix A), we address the static question. That is, we examine conditions that allow us (in the most general circumstances possible) to characterize optima. We use these conditions to test candidate points to see whether they are (or are not) minima, maxima, saddlepoints, or none of the above. In doing so, we are interested not in the generation of the candidate points but in determining if they solve the unconstrained multivariable problem: Minimize ƒ(x)

x 僆 RN

(3.1)

in the absence of constraints on x, where x is a vector of design variables of dimension N and ƒ is a scalar objective function. We typically assume that xi (for all i ⫽ 1, 2, 3, . . . , N) can take on any value, even though in many practical applications x must be selected from a discrete set. In addition, we shall often find it convenient to assume that ƒ and its derivatives exist and are continuous everywhere, even though we know that optima may occur at points of discontinuity in ƒ or its gradient, ⵜƒ ⫽

78

冋

册

⭸ƒ ⭸ƒ ⭸ƒ ⭸ƒ , , ,..., ⭸x1 ⭸x2 ⭸x3 ⭸xN

T

(3.2)


FUNCTIONS OF SEVERAL VARIABLES

79

It is well to remember that ƒ may take on a minimum value at a point x, where ƒ is discontinuous or ⵜƒ is discontinuous or does not exist. We must at least temporarily discard these complicating factors in order to develop useful optimality criteria. Finally, we must often be satisfied to identify local optima, since the nonlinear objective ƒ will typically not be convex (see Appendix B) and therefore will be multimodal, as in Figure 3.1, a contour plot of Himmelblau’s function: ƒ(x) ⫽ [x 21 ⫹ x2 ⫺ 11]2 ⫹ [x1 ⫹ x 22 ⫺ 7]2

(3.3)

We see that this function has four distinct minima. Next, we examine the dynamic question, given x (0), a point that does not satisfy the above-mentioned optimality criteria, what is a better estimate x (1) of the solution x*? This leads naturally to a discussion of a number of methods, which constitutes the greater part of the chapter. We classify the methods according to their need for derivative information. The chapter closes with a

Figure 3.1. Contours of a multimodal function.

80


discussion of the relative merits of the various methods with supportive numerical results.

3.1

OPTIMALITY CRITERIA

Here we develop conditions or tests that will allow us to characterize (that is, classify) points in the design space. We examine optimality criteria for basically two reasons: (1) because they are necessary to recognize solutions and (2) because they provide motivation for most of the useful methods. Consider the Taylor expansion of a function of several variables: ƒ(x) ⫽ ƒ(x) ⫹ ⵜƒ(x)T ⌬ x ⫹ –21 ⌬ xT ⵜ2ƒ(x) ⌬ x ⫹ O3(⌬ x)

(3.4)

where x ⫽ current or expansion point in RN ⌬ x ⫽ x ⫺ x ⫽ change in x ⵜƒ(x) ⫽ N-component column vector of first derivatives of ƒ(x) evaluated at x ⵜ2ƒ(x) ⫽ Hƒ(x) ⫽ N ⫻ N symmetric matrix of second partial derivatives of ƒ(x) evaluate at x, often called the Hessian matrix. The element in the ith row and jth column is ⭸2ƒ/ ⭸xi ⭸xj. O3(⌬ x) ⫽ all terms of order greater than 2 in ⌬ x Let us now ignore the higher order terms [that is, drop O3(⌬ x)] and examine the change in the magnitude of the objective ƒ(x) corresponding to arbitrary changes in x. ⌬ƒ(x) ⫽ ƒ(x) ⫺ ƒ(x) ⫽ ⵜƒ(x)T ⌬ x ⫹ –12 ⌬ xT ⵜ2ƒ(x) ⌬ x

(3.5)

Recall that by definition a minimum is a point such that all other points in a neighborhood produce a greater objective value. That is, ⌬ƒ ⫽ ƒ(x) ⫺ ƒ(x) ⭓ 0

(3.6)

The point x is a global minimum if (3.6) holds for all x 僆 RN, and we give this point the symbol x**. When (3.6) holds for some ␦-neighborhood, that is, for all x such that 㛳 x ⫺ x㛳 ⭐ ␦ for some ␦ ⬎ 0, then x is a local minimum or x*. When ⌬ƒ ⫽ ƒ(x) ⫺ ƒ(x) ⭐ 0

(3.7)

then x is a maximum, either local or global as before. Removal of the equality in (3.6) and (3.7) produces strict minimum and maximum points. When ⌬ƒ

3.1

OPTIMALITY CRITERIA

81

is either positive, negative, or zero depending on the choice of nearby points in a ␦-neighborhood, then x is a saddlepoint. Let us return to reason on (3.5), and recall that we assume that ƒ(x), ⵜƒ(x), and ⵜ2ƒ(x) exist and are continuous for all x 僆 RN. Therefore, from (3.5), in order that the sign of ⌬ƒ be known for arbitrary values of ⌬ x, ⵜƒ(x) must be zero; that is, x must be a stationary point. Otherwise, we could force ⌬ƒ to be plus or minus depending on the sign of ⵜƒ(x) and ⌬ x. Accordingly, x must satisfy the stationary conditions: ⵜƒ(x) ⫽ 0

(3.8)

⌬ƒ(x) ⫽ –12 ⌬ xT ⵜ2ƒ(x) ⌬ x

(3.9)

so (3.5) becomes

Clearly, then, the sign of ⌬ƒ(x) depends on the nature of the quadrative form; Q(x) ⫽ ⌬ xT ⵜ2ƒ(x) ⌬ x

(3.10)

or Q(z) ⫽ zTAz We know from linear algebra (see Appendix A) that A is positive definite if

for all z, Q(z) ⬎ 0

A is positive semidefinite if

for all z, Q(z) ⭓ 0

A is negative definite if

for all z, Q(z) ⬍ 0

A is negative semidefinite if

for all z, Q(z) ⭐ 0

A is indefinite if

for some z, Q(z) ⬎ 0

and

for other z, Q(z) ⬍ 0

(3.11)

Therefore from (3.11) the stationary point x is a Minimum if ⵜ2ƒ(x) is positive definite Maximum if ⵜ2ƒ(x) is negative definite

(3.12)

Saddlepoint if ⵜ2ƒ(x) is indefinite In addition, it may be helpful to examine the nature of the stationary point x in a slightly different light. Consider the stationary point x, a surrounding ␦-neighborhood, and directions emanating from x (Figure 3.2) such that

82


Figure 3.2. A ‘‘stationary’’ neighborhood.

x˜ ⫽ x ⫹ ␣s(x)

(3.13)

We can generate all points in the neighborhood x˜ by appropriate choice of ␣ and s. Substitution of (3.13) into (3.9) produces ⌬ƒ(x) ⫽

␣2 T 2 s ⵜ ƒ(x)s 2

(3.14)

Now we can classify s(x) as a descent direction, ascent direction, or neither using the previous reasoning in (3.11) and (3.12). Hence x becomes x*, a local minimum when a descent direction cannot be found, and this corresponds to the case when ⵜ2ƒ(x) is positive semidefinite. We are now in a position to state the necessary and sufficient conditions for the existence of a minimum of a function of several variables. Theorem 3.1

Necessary Conditions

For x* to be a local minimum, it is necessary that ⵜƒ(x*) ⫽ 0

(3.15a)

ⵜ2ƒ(x*) is positive semidefinite

(3.15b)

and

3.1

Theorem 3.2

OPTIMALITY CRITERIA

83

Sufficient Conditions

If ⵜƒ(x*) ⫽ 0

(3.16a)

ⵜ2ƒ(x*) is positive definite

(3.16b)

and

then x* is an isolated local minimum of ƒ(x). The proofs of these theorems follow directly from the previous discussion and are left as exercises for the reader. Typically we must be satisfied to find a local minimum, but when we can show that xT ⵜ2ƒ(x)x ⭓ 0 for all x, we say that ƒ(x) is a convex function and a local minimum is a global minimum. Example 3.1

Optimality Criteria

Consider the function ƒ(x) ⫽ 2x 21 ⫹ 4x1x 32 ⫺ 10x1x2 ⫹ x 22 as shown in Figure 3.3, and the point x ⫽ [0, 0]T. Classify x.

Figure 3.3. Two-variable nonlinear function of Example 3.1.

84


Solution ⭸ƒ ⫽ 4x1 ⫹ 4x 32 ⫺ 10x2 ⭸x1 ⭸ƒ ⫽ 12x1x 22 ⫺ 10x1 ⫹ 2x2 ⭸x2 ⵜƒ(x) ⫽ [0, 0]T

Therefore, x is a stationary point. ⭸2ƒ ⫽4 ⭸x 21 ⭸2ƒ ⫽ 24x1x2 ⫹ 2 ⫽ ⫹2 ⭸x 22 ⭸2ƒ ⫽ 12x 22 ⫺ 10 ⫽ ⫺10 ⭸x1⭸x2

Therefore, ⵜ2ƒ(x) ⫽ Hƒ(x) ⫽

冋

⫹4 ⫺10

⫺10 ⫹2

册

ⵜ2ƒ(x) is indefinite, since the quadratic form zTHƒ z is greater than zero for z ⫽ (0, 1) and less than zero for z ⫽ (1, 1) (see Appendix A). Accordingly, x

is a saddlepoint, as the contour map (Figure 3.3) suggests. 3.2

DIRECT-SEARCH METHODS

In this and the following sections of this chapter we consider the dynamic question for functions of several variables. That is, we examine methods or algorithms that iteratively produce estimates of x*, that set of design variables that causes ƒ(x) to take on its minimum value. The methods work equally well for maximization by letting the objectives be ⫺ƒ(x). The methods that have been devised for the solution of this problem can be classified into three broad categories based on the type of information that must be supplied by the user: 1. Direct-search methods, which use only function values 2. Gradient methods, which require estimates of the first derivative of ƒ(x) 3. Second-order methods, which require estimates of the first and second derivatives of ƒ(x)

3.2


85

We will consider examples from each class, since no one method or class of methods can be expected to uniformly solve all problems with equal efficiency. For instance, in some applications available computer storage is limited; in others, function evaluations are very time consuming; in still others, great accuracy in the final solution is desired. In some applications it is either impossible or else very time consuming to obtain analytical expressions for derivatives. In addition, analytical derivative development is prone to error. Consequently, different approximations must be employed if gradient-based techniques are to be used. This in turn may mean considerable experimentation to determine step sizes that strike the proper balance between roundoff and truncation errors. Clearly it behooves the engineer to tailor the method used to the characteristics of the problem at hand. Methods for the unconstrained minimization problem are relatively well developed compared to the other areas of nonlinear programming. In fact, excellent surveys of the most powerful method do exist [1–3]. The books by Murray [1] and Fletcher [3] are good examples. In this section we treat direct methods, which are methods that require only values of the objective to proceed, and in the following section we discuss gradient and second-order methods. We assume here that ƒ(x) is continuous and ⵜƒ(x) may or may not exist but certainly is not available. In addition, we recognize that these direct methods can be used on problems where ⵜƒ does exist and often are when ⵜƒ is a complex vector function of the design variables. Finally, in this and the following sections we assume that ƒ(x) has a single minimum in the domain of interest. When these methods are applied to multimodal functions, we must be content to locate local minima. Direct-search methods are especially useful when there is ‘‘noise’’ or uncertainty in the objective function evaluation. Multivariable methods that employ only function values to guide the search for the optimum can be classified broadly into heuristic techniques and theoretically based techniques. The heuristic techniques, as the name implies, are search methods constructed from geometric intuition for which no performance guarantees other than empirical results can be stated. The theoretically based techniques, on the other hand, do have a mathematical foundation that allows performance guarantees, such as convergence, to be established, at least under restricted conditions. We will examine in particular detail three direct search techniques: 1. Simplex search, or S 2, method 2. Hooke–Jeeves pattern search method 3. Powell’s conjugate direction method The first two of these are heuristic techniques and exhibit fundamentally different strategies. The S 2 method employs a regular pattern of points in the design space sequentially, whereas the Hooke–Jeeves method employs a fixed set of directions (the coordinate directions) in a recursive manner. Powell’s method is a theoretically based method that was devised assuming a quadratic

86


objective and for such functions will converge in a finite number of iterations. The common characteristic of the three methods is that computationally they are relatively uncomplicated, hence easy to implement and quick to debug. On the other hand, they can be, and often are, slower than the derivativebased methods. We give little attention to methods based upon the region elimination concepts of Chapter 2, such as given by Box et al. [4] and Krolak and Cooper [5], since they tend to be vastly inferior to other available methods. 3.2.1

The S 2 (Simplex Search) Method

The earliest attempts to solve unconstrained problems by direct search produced methods modeled essentially on single-variable methods. Typically, the domain of the performance index was subdivided into a grid of points, and then various strategies for shrinking the area in which the solution lay were applied. Often this led to nearly exhaustive enumeration and hence proved unsatisfactory for all but two-dimensional problems. A more useful idea was to pick a base point and, rather than attempting to cover the entire range of the variables, evaluate the performance index in some pattern about the base point. For example, in two dimensions, a square pattern such as in Figure 3.4 is located. Then the best of the five points is selected as the next base point around which to locate the next pattern of points. If none of the corner points is better than the base point, the scale of the grid is reduced and the search continues. This type of evolutionary optimization was advocated by Box [6] and others to optimize the performance of operating plants when there is error present in the measured response of the plant to imposed process variable changes. In higher dimensions this corresponds to evaluating the performance index at each vertex as well as the centroid of a hypercube and is called a factorial design pattern. If the number of variables, that is, the dimension of the search space, is N, then a factorial search pattern will require 2N ⫹ 1 function evaluations per pattern. The number of evaluations, even for problems of modest dimension, rapidly becomes overwhelming. Hence, even though the logic here is simple, more efficient methods must be investigated if we are to use these direct-search methods on problems of reasonable size.

Figure 3.4. Factorial design.

3.2


87

One particularly ingenious search strategy is the simplex search method of Spendley, Hext, and Himsworth [7]. This method and others in this class have no relationship to the simplex method of linear programming. The similarity in name is indeed unfortunate. The simplex search method of Spendley, Hext, and Himsworth is based upon the observation that the first-order experimental design requiring the smallest number of points is the regular simplex. In N dimensions, a regular simplex is a polyhedron composed of N ⫹ 1 equidistant points, which form its vertices. For example, an equilateral triangle is a simplex in two dimensions; a tetrahedron is a simplex in three dimensions. The main property of a simplex employed by their algorithm is that a new simplex can be generated on any face of the old one by projecting any chosen vertex a suitable distance through the centroid of the remaining vertices of the old simplex. The new simplex is then formed by replacing the old vertex by the newly generated projected point. In this way each new simplex is generated with a single evaluation of the objective. This process is demonstrated for two dimensions in Figure 3.5. The method begins by setting up a regular simplex in the space of the independent variables and evaluating the function at each vertex. The vertex with highest functional value is located. This ‘‘worst’’ vertex is then reflected through the centroid to generate a new point, which is used to complete the next simplex. As long as the performance index decreases smoothly, the iterations move along crabwise until either the minimum is straddled or the iterations begin to cycle between two or more simplexes. These situations are resolved using the following three rules: Rule 1

Minimum ‘‘Straddled’’

If the selected ‘‘worse’’ vertex was generated in the previous iteration, then choose instead the vertex with the next highest function value. Rule 2

Cycling

If a given vertex remains unchanged for more than M iterations, reduce the size of the simplex by some factor. Set up a new simplex with the currently

Figure 3.5. Construction of new simplex.

88


lowest point as the base point. Spendley et al. suggest that M be predicted via M ⫽ 1.65N ⫹ 0.05N 2 where N is the problem dimension and M is rounded to the nearest integer. This rule requires the specification of a reduction factor. Rule 3

Termination Criterion

The search is terminated when the simplex gets small enough or else if the standard deviation of the function values at the vertices gets small enough. This rule requires the specification of a termination parameter. The implementation of this algorithm requires only two types of calculations: (1) generation of a regular simplex given a base point and appropriate scale factor and (2) calculation of the reflected point. The first of these calculations is readily carried out, since it can be shown from elementary geometry that given an N-dimensional starting or base point x (0) and a scale factor ␣, the other N vertices of the simplex in N dimensions are given by

再

x (0) ⫹ ␦1 j (0) x j ⫹ ␦2

x j(i) ⫽

if j ⫽ i if j ⫽ i

(3.17)

for i and j ⫽ 1, 2, 3, . . . , N. The increments ␦1 and ␦2, which depend only on N and the selected scale factor ␣, are calculated from ␦1 ⫽ ␦2 ⫽

冋冋

册

(N ⫹ 1)1 / 2 ⫹ N ⫺ 1 ␣ N兹2

(3.18)

(N ⫹ 1)1 / 2 ⫺ 1 ␣ N兹2

(3.19)

册

Note that the scale factor ␣ is chosen by the user to suit the problem at hand. The choice ␣ ⫽ 1 leads to a regular simplex with sides of unit length. The second calculation, reflection through the centroid, is equally straightforward. Suppose x ( j) is the point to be reflected. Then the centroid of the remaining N points is xc ⫽

1 N

冘x N

(i)

i⫽0 i⫽j

All points on the line from x ( j) through xc are given by

(3.20)

3.2


x ⫽ x ( j) ⫹ ␭(xc ⫺ x ( j))

89

(3.21)

The choice ␭ ⫽ 0 yields the original point x ( j), while the choice ␭ ⫽ 1 corresponds to the centroid xc. To retain the regularity of the simplex, the reflection should be symmetric. Hence, ␭ ⫽ 2 will yield the desired new vertex point. Thus, j) j) x (new ⫽ 2xc ⫺ x (old

(3.22)

Both calculations are illustrated in the following example. Example 3.2

Simplex Search Calculations

Minimize ƒ(x) ⫽ (1 ⫺ x1)2 ⫹ (2 ⫺ x2)2 Solution. The construction of the initial simplex requires the specification of an initial point and a scale factor. Suppose x (0) ⫽ [0, 0]T and ␣ ⫽ 2 are selected. Then ␦1 ⫽

冋

2兹2

␦2 ⫽

冋

2兹2

兹3 ⫹ 1

册

␣ ⫽ 1.9318

and 兹3 ⫺ 1

册

␣ ⫽ 0.5176

With these two parameters, the other two vertices are calculated as x (2) ⫽ [0 ⫹ 0.5176, 0 ⫹ 1.9318]T ⫽ [0.5176, 1.9318]T x (1) ⫽ [0 ⫹ 1.9318, 0 ⫹ 0.5176]T ⫽ [1.9318, 0.5176]T with function values ƒ(x (1)) ⫽ 3.0658 and ƒ(x (2)) ⫽ 0.2374. Since ƒ(x (0)) ⫽ 5, x (0) is the point to be reflected to form the new simplex. The replacement x (3) is calculated as follows: xc ⫽

1 2

冘x 2

i⫽1

Therefore, using (3.22) we get

(i)

⫽

1 (1) (x ⫹ x (2)) 2

90


x (3) ⫽ x (1) ⫹ x (2) ⫺ x (0) x (3) ⫽ [2.4494, 2.4494]T At the new point ƒ(x (3)) ⫽ 2.3027, an improvement. The new simplex is composed of points x (1), x (2), and x (3). The algorithm would now continue by reflecting out the point with the highest function value, x (1). The iteration proceeds as before except when situations are encountered that require rules 1, 2, or 3 given before. The S 2 algorithm as stated above has several obvious advantages: 1. The calculations are simple and the logic uncomplicated, hence a program will be short. 2. Storage requirements are relatively low: essentially an array of dimension (N ⫹ 1, N ⫹ 2). 3. Comparatively few adjustable parameters need to be supplied: the scale factor ␣, a factor by which to reduce ␣ if rule 2 is encountered, and termination parameters. 4. The algorithm is effective when evaluation errors are significant because it operates on the worst rather than the best point. These factors make it quite useful for on-line performance improvement calculations. The algorithm also has several serious disadvantages: 1. It can encounter scaling problems because all coordinates are subjected to the same scale factor ␣. To mitigate this, the user should in any given application scale all variables to have comparable magnitude. 2. The algorithm is slow, since it does not use past information to generate acceleration moves. 3. There is no simple way of expanding the simplex without recalculating the entire pattern. Thus, once ␣ is reduced (because, e.g., a region of narrow contours is encountered), the search must continue with this reduced step size. To partially eliminate some of these disadvantages, a modified simplex procedure has been developed by Nelder and Mead [8]. They observed that although it is convenient to use the formula for a regular simplex when generating the initial pattern, there is no real incentive for maintaining the regularity of the simplex as the search proceeds. Consequently, they allow both expansion and contraction in the course of the reflection calculation. Their modifications require consideration of the point with highest current function value, x (h); the next highest point, x ( g); the point with the lowest current func-

3.2


91

tion value, x (l); and the corresponding function values ƒ(h), ƒ( g), and ƒ(l). Recall that the reflection step is described by the line x ⫽ x (h) ⫹ ␭(xc ⫺ x (h)) or x ⫽ x (h) ⫹ (1 ⫹ ␪)(xc ⫺ x (h))

(3.23)

If we let ␪ ⫽ 1, we get the normal simplex reflection, such that xnew is located a distance 㛳 xc ⫺ x ( j)㛳 from xc. When ⫺1 ⭐ ␪ ⬍ 1, we produce a shortened reflection, or a contraction, while the choice ␪ ⬎ 1 will generate a lengthened reflection step, or an expansion of the simplex. The situations are depicted in Figure 3.6. The three values of ␪ used for normal reflection, contraction, and expansion are denoted respectively ␣, ␤, and ␥. The method proceeds from the initial simplex by determination of x (h), x ( g), x (l), and xc. A normal reflection of the simplex function values is checked to see if termination is appropriate. If not, a normal reflection, expansion, or contraction is taken using the tests outlined in Figure 3.6. The iterations continue until the simplex function values do not vary significantly. Nelder and Mead recommend that the values ␣ ⫽ 1, ␤ ⫽ 0.5, and ␥ ⫽ 2 be employed. Some limited numerical comparisons indicate that the Nelder–Mead variant is very reliable in the presence of noise or error in the objective and is

Figure 3.6. Expansion and contraction of a simplex.

92


reasonably efficient. In 1969, Box and Draper [9] stated that this algorithm is the ‘‘most efficient of all current sequential techniques.’’ A 1972 study by Parkinson and Hutchinson [10] investigated the effects of the choice of ␣, ␤, ␥, and the manner of construction of the initial simplex. They determined that the shape of the initial simplex was not important but its orientation was. The best overall choice of parameters was (␣, ␤, ␥) ⫽ (2, 0.25, 2.5) and this choice worked well if repeated successive expansions were allowed. 3.2.2

Hooke–Jeeves Pattern Search Method

The preceding search techniques were based on the systematic disposition and manipulation of a pattern of trial points. Yet despite our preoccupation with the geometric configuration in which the trial points are located, it is apparent that the main role of the set of points is to generate a direction of search. The exact placement of the points really only influences how sensitive the generated search direction is to local variations in the function topology. In particular, the equation for the reflection step, xnew ⫽ x ( j) ⫹ ␭(xc ⫺ x ( j)) clearly reveals that the entire set of points is condensed into a single vector difference that defines a direction. The remaining search logic is merely concerned with adjusting the step size ␭ so that reasonable improvement is attained. If the generation of search directions is the primary motivation for manipulating the pattern of trial points, then a reasonable improvement over the simplex strategy might be to operate directly with a set of direction vectors using them to guide the search. In the simplest approach, a fixed arbitrary set of directions could be selected and searched recursively for improvement. Alternatively, one could devise strategies for modifying one or more of the search directions with each iteration so as to better align them with the gross topology of the function. In either case, to ensure that the complete domain of the function is searched, it is reasonable to insist that the set of search directions be independent and span the entire domain of ƒ(x). For example, it clearly would be inadequate to search for the optimum of a three-variable function by alternating use of two search directions. Hence, all the directsearch methods discussed here will require at least N independent search directions, where N is the dimension of x. An elementary example of a method that recursively uses a fixed set of search directions is the ‘‘one-variable-at-a-time,’’ or sectioning, technique. In this approach the set of directions is chosen to be the coordinate directions in the space of the problem variables. Each of the coordinate directions is sequentially searched using a single-variable optimization scheme until no further improvement is possible. If the objective function is spherically symmetrical, then this search will succeed admirably. However, if the function contours are distorted or elongated, as is usually the case, then the iteration

3.2


93

can degenerate into an infinite sequence of shorter and shorter steps, and the search becomes quite inefficient. In fact, it has been shown by Powell [11] that the cyclic use of coordinate searches (or in fact of any fixed set of search directions) can not only prove inefficient but actually fail to converge to a local optimum even if an infinite number of iterations are employed. In the course of attempting to improve the efficiency of this method, it was noted that considerable acceleration in convergence could be achieved by periodically searching in the direction d (i) ⫽ x (i) ⫺ x (i⫺N). This simple observation forms the basis of a modified sectioning method devised by Hooke and Jeeves [12], one of the first algorithms to incorporate the past history of a sequence of iterations into the generation of a new search direction. Basically, the Hooke–Jeeves (HJ) procedure is a combination of ‘‘exploratory’’ moves of the one-variable-at-a-time kind with ‘‘pattern’’ or acceleration moves regulated by some heuristic rules. The exploratory moves examine the local behavior of the function and seek to locate the direction of any sloping valleys that might be present. The pattern moves utilize the information generated in the exploration to step rapidly along the valleys. Exploratory Moves. Given a specified step size, which may be different for

each coordinate direction and change during the search, the exploration proceeds from an initial point by the specified step size in each coordinate direction. If the function value does not increase, the step is considered successful. Otherwise, the step is retracted and replaced by a step in the opposite direction, which in turn is retained depending upon whether it succeeds or fails. When all N coordinates have been investigated, the exploratory move is completed. The resulting point is termed a base point. Pattern Move. A pattern move consists of a single step from the present base point along the line from the previous to the current base point. Thus, a new pattern point is calculated as

x (k⫹1) ⫽ x (k) ⫹ (x (k) ⫺ x (k⫺1)) p

(3.24)

Now, recognizing that this move may not result in an improvement, the point x (k⫹1) is accepted only temporarily. It becomes the temporary base point for p a new exploratory move. If the result of this exploratory move is a better point than the previous base point x (k), then this point is accepted as the new base point x (k⫹1). On the other hand, if the exploratory move does not produce improvement, then the pattern move is discarded and the search returns to x (k), where an exploratory search is undertaken to find a new pattern. Eventually a situation is reached when even this exploratory search fails. In this case the step sizes are reduced by some factor and the exploration resumed. The search is terminated when the step size becomes sufficiently small. To summarize the logic of the method, we employ the following notation:

94


x (k) x (k⫺1) x (k⫹1) p x (k⫹1)

Current base point Previous base point Pattern move point Next (or new) base point

Using this notation, we outline the pattern search method of Hooke and Jeeves. Pattern Search Method of Hooke and Jeeves

Step 1. Define: The starting point x (0) The increments ⌬i for k ⫽ 1, 2, 3, . . . , N The step reduction factor ␣ ⬎ 1 A termination parameter ␧ ⬎ 0 Step 2. Perform exploratory search. Step 3. Was exploratory search successful (i.e., was a lower point found)? Yes: Go to 5. No: Continue. Step 4. Check for termination. Is 㛳⌬㛳 ⬍ ␧? Yes: Stop; current point approximates x*. No: Reduce the increments: ⌬i ⫽

⌬i ␣

i ⫽ 1, 2, 3, . . . , N

Go to 2. Step 5. Perform pattern move: x (k⫹1) ⫽ x (k) ⫹ (x (k) ⫺ x (k⫺1)) p Step 6. Perform exploratory research using x (k⫹1) as the base point; let the p result be x (k⫹1). Step 7. Is ƒ(x (k⫹1)) ⬍ ƒ(x (k))? Yes: Set x (k⫺1) ⫽ x (k); x (k) ⫽ (k⫹1). Go to 5. No: Go to 4. Example 3.3

Hooke–Jeeves Pattern Search

Find the minimum of ƒ(x) ⫽ 8x 21 ⫹ 4x1x2 ⫹ 5x 22 from x (0) ⫽ [⫺4, ⫺4]T.

3.2


95

Solution. To use the HJ direct-search method we must choose ⌬ ⫽ step-size increments ⫽ [1, 1]T

␣ ⫽ step reduction factor ⫽ 2 ␧ ⫽ termination parameter ⫽ 10⫺4

Begin the iteration with an exploration about x (0), with ƒ(x (0)) ⫽ 272. With x2 fixed, we increment x1: With x2 ⫽ ⫺4 x1 ⫽ ⫺4 ⫹ 1 → ƒ(⫺3, ⫺4) ⫽ 200 ⬍ ƒ(x (0)) → success Therefore, we fix x1 at ⫺3 and increment x2: With x1 ⫽ ⫺3 x2 ⫽ ⫺4 ⫹ 1 → ƒ(⫺3, ⫺3) ⫽ 153 ⬍ 20 → success Therefore the result of the first exploration is x (1) ⫽ [⫺3, ⫺3]T

ƒ(x (1)) ⫽ 153

Since the exploration was a success, we go directly to the pattern move: x 2p ⫽ x (1) ⫹ (x (1) ⫺ x (0)) ⫽ [⫺2, ⫺2]T ƒ(x p(2)) ⫽ 68 Now we perform an exploratory search about x (2) p . We find that positive increments on x1 and x2 produce success. The result is the point x (2) ⫽ [⫺1, ⫺1]T

ƒ(x (2)) ⫽ 17

Since ƒ(x (2)) ⬍ ƒ(x (1)), the pattern move is deemed a success and x (2) becomes the new base point for an additional pattern move. The method continues until step reduction causes termination near the solution, x* ⫽ [0, 0]T. Progress of the method is displayed in Figure 3.7. As can be seen from Example 3.3, the search strategy is quite simple, the calculations uninvolved, and the storage requirements less than even that associated with the simplex search method. Because of these factors, the HJ algorithm has enjoyed wide application in all areas of engineering, especially when used with penalty functions (to be discussed in Chapter 6). Because of its reliance on coordinate steps, the algorithm can, however, terminate prematurely, and in the presence of severe nonlinearities will degenerate to a sequence of exploratory moves without benefit of pattern acceleration. Numerous modifications to the basic HJ method have been reported since it was introduced. For instance, Bandler and McDonald [13] modified the

96


Figure 3.7. Pattern search iterations for Example 3.3.

basic HJ procedure by inserting rules for expanding and contracting the exploratory increments as well as allowing a contracted pattern step if the normal pattern step fails. Our own experiments suggest the addition of another phase of the algorithm, which we call pattern exploitation. That is, when a pattern move is a success, why not completely exploit the direction by conducting a line search along the pattern direction or at least trying pattern steps of increasing magnitude? This can often significantly accelerate the convergence of the method. Emery and O’Hagan [14] altered the exploratory phase by using a set of orthogonal search directions whose orientation was redirected randomly after each iteration. Rosenbrock [15] developed a method that, like HJ, generated a new search direction based on the accumulated progress of recent iterations. However, unlike HJ, Rosenbrock’s approach continually changed the set of direction vectors used in the exploratory phase by a process of orthogonalizations. Another technique, reported by Swann [16] and sometimes called the DSC (Davidon–Swann–Campey) multivariate search method, used a strategy similar to Rosenbrock’s. Rather than relying on only simple increments for each direction, a complete line search is performed in each direction. Each of the many variants claims to offer advantages

3.2


97

over its class siblings in selected applications. However, it is questionable whether the added complications are justifiable. If a more sophisticated algorithm is in fact to be used, then a better algorithm than these heuristic methods is available, namely, that due to Powell. 3.2.3

Powell’s Conjugate Direction Method

To date the most successful of the direct-search algorithms is the method due to Powell [17], especially with the modifications suggested by Zangwill [18] and Brent [19]. This algorithm effectively uses the history of the iterations to build up directions for acceleration and at the same time avoids degenerating to a sequence of coordinate searches. It is based upon the model of a quadratic function and thus has a theoretical basis. There are two reasons for choosing a quadratic model: 1. It is the simplest type of nonlinear function to minimize (linear functions do not take on interior optima), and hence any general technique must work well on a quadratic if it is to have any success with a general function. 2. Near the optimum, all nonlinear functions can be approximated by a quadratic (since in the Taylor expansion the linear part must vanish). Hence, the behavior of the algorithm on the quadratic will give some indication of how the algorithm will converge for general functions. The motivation for the algorithm stems from the observation that if a quadratic function in N variables can be transformed so that it is just the sum of perfect squares, then the optimum can be found after exactly N single-variable searches, one with respect to each of the transformed variables. The process of transforming a quadratic function q(x) ⫽ a ⫹ bTx ⫹ –12 xTCx

(3.25)

into a sum of perfect squares is equivalent to finding a transformation matrix T such that the quadratic term is reduced to diagonal form. Thus, given the quadratic form Q(x) ⫽ xTCx

(3.26)

x ⫽ Tz

(3.27)

the desired transformation

will yield

98


Q(x) ⫽ zTTTCTz ⫽ zTDz

(3.28)

where D is a diagonal matrix, that is, its elements are nonzero only if i ⫽ j. Let tj be the jth column of T. Then the transformation in (3.27) expresses the fact that we are rewriting each vector x as a linear combination of the column vectors tj: x ⫽ Tz ⫽ t1z1 ⫹ t2z2 ⫹ 䡠䡠䡠 ⫹ tNzN

(3.29)

In other words, instead of writing x in terms of the standard coordinate system represented by the set of vectors e(i), we are expressing it in terms of a new coordinate system given by the set of vectors tj. Moreover, since they diagonalize the quadratic, the set of vectors tj correspond to the principal axes of the quadratic form. Graphically, this corresponds to taking a general quadratic function with cross terms, as shown in Figure 3.8, and realigning the new coordinate axes to coincide with the major and minor axes of the quadratic, as shown in Figure 3.9. To summarize, by taking the quadratic through the transformation, we are really choosing a new coordinate system for the quadratic that coincides with

Figure 3.8. Quadratic with cross terms.

3.2


99

Figure 3.9. A quadratic without cross terms.

the principal axes of the quadratic. Consequently, single-variable searches for the optimum performed in the space of the transformed or z variables simply correspond to single-variable searches along each of the principal axes of the quadratic. Since the principal axes are the same as the vectors tj, the singlevariable searches are actually performed along each of these vectors. Let us illustrate with an example. Example 3.4

Sum-of-Squares Transformation

Consider the function ƒ(x) ⫽ 4x 21 ⫹ 3x 22 ⫺ 4x1x2 ⫹ x1 and the transformation x1 ⫽ z1 ⫹ –21 z2 or

x2 ⫽ z2

100


x⫽

冋册 1 0

–12

1

z

The transformed quadratic becomes ƒ(z) ⫽ 4z 21 ⫹ 2z 22 ⫹ z1 ⫹ –12 z2 Note that this transformation is not unique, since the tj’s are not scaled to be orthonormal. For instance, you can readily verify that the transformation x⫽

冋册 1 –23

0 z 1

will also diagonalize the quadratic function. Given x (0) ⫽ [0, 0]T and the two columns of the transformation matrix, t1 ⫽ [1, 0]T

t2 ⫽ [–12 , 1]T

3 1 T the optimum [⫺–– 16 , ⫺– 8 ] should be found by two successive single-variable searches in the t1 and t2 directions. Searching along t1 first,

x (1) ⫽ x (0) ⫹ ␭t1 we find ␭ ⫽ ⫺–18 and x (1) ⫽ [⫺–18 , 0]T. Now perform the second search along 3 1 T t2 from x (1). Again, ␭ ⫽ ⫺–18 and x (2) ⫽ [⫺–– 16 , ⫺– 8 ] , the solution. Generalizing from the above example and the preceding analysis, if a suitable set of transforming vectors tj, j ⫽ 1, . . . , N, conventionally called conjugate directions, can be obtained, then the optimum of a quadratic function can be found by exactly N single-variable searches, one along each of the N directions tj, j ⫽ 1, . . . , N. The question thus remains how to calculate a set of such vectors. Clearly, if an estimate of the matrix C were at hand, then the transformation T could be obtained by Gaussian elimination (as explained in Appendix A) followed by a matrix inversion. Gaussian elimination will yield the factorization C ⫽ PTDP

(3.30)

hence, (P⫺1)TC(P⫺1) ⫽ D

and

T ⫽ P⫺1

(3.31)

would satisfy our needs. However, an estimate of C is not available in our case, because we are seeking to develop a method for optimizing ƒ(x) that

3.2


101

uses only function values, not first derivatives and certainly not second derivatives. Fortunately, a set of conjugate directions can nonetheless be obtained by using only function values on the basis of the following elementary property of quadratic functions. Parallel Subspace Property Given a quadratic function q(x), two arbitrary but distinct points x (1) and x (2), and a direction d, if y (1) is the solution to min q(x (1) ⫹ ␭d) and y (2) is the solution to min q(x (2) ⫹ ␭d), then the direction y (2) –y (1) is C conjugate to d. In two dimensions, this property is illustrated in Figure 3.10, where we see that a single-variable search from y (1) or y (2) along the direction y (2) –y (1) will produce the minimum. Thus, it can be concluded that in two dimensions, by performing three single-variable searches, a set of conjugate directions can be generated, and in addition the optimum of the quadratic can be obtained. Before proceeding to exploit this construction further, we pause to develop the proof of the above very important property. First, recall that by definition a set of C conjugate directions is a set of column vectors with the property that the matrix T formed from these columns diagonalizes C: TTCT ⫽ D

(3.32)

Alternatively, since all off-diagonal elements of D are zero, this means that

Figure 3.10. Conjugacy in two dimensions.

102


tTj Ctk ⫽

再

dkk 0

if j ⫽ k if j ⫽ k

(3.33)

where ti is the ith column of T. Hence, we are able now to give the more conventional, equivalent, and possibly more useful definition of conjugacy. Conjugate Directions

Given an N ⫻ N symmetric matrix C, the directions s (1), s (2), s (3), . . . , s (r), r ⭐ N, are said to be C conjugate if the directions are linearly independent, and s (i)TCs ( j) ⫽ 0

for all i ⫽ j

(3.34)

Consider once again the general quadratic function given earlier: q(x) ⫽ a ⫹ bTx ⫹ –12 xTCx The points along the direction d from x (1) are x ⫽ x (1) ⫹ ␭d and the minimum of q(x) along d is obtained by finding ␭* such that ⭸q/ ⭸␭ ⫽ 0. The derivative is calculated using the chain rule: ⭸q ⭸q ⭸x ⫽ ⫽ (bT ⫹ xTC) d ⭸␭ ⭸x ⭸␭

(3.35)

By hypothesis the minimum occurs at y (1); hence, [(y (1))TC ⫹ bT] d ⫽ 0

(3.36)

Similarly, since the minimum of q(x) along d from x (2) is attained at y (2), we have [(y (2))TC ⫹ bT] d ⫽ 0

(3.37)

Subtracting Eq. (3.36) from Eq. (3.37), the result is (y (2) ⫺ y (1))TCd ⫽ 0

(3.38)

Accordingly, by definition, the directions d and y (2) –y (1) are C conjugate, and the parallel subspace property of quadratic functions has been verified.

3.2

Example 3.5


103

Minimization via the Parallel Subspace Property

We illustrate the above property by considering once again the quadratic q(x) ⫽ 4x 21 ⫹ 3x 22 ⫺ 4x1x2 ⫹ x1 Suppose the two points x (1) ⫽ [0, 0]T, x (2) ⫽ [1, 0]T, and the direction d ⫽ [1, 1]T are given. The first search follows the line x ⫽ [0, 0]T ⫹ ␭[1, 1]T and yields the point (␭* ⫽ ⫺–16 ); y (1) ⫽ [⫺–16 , ⫺–16 ]T. The second search follows the line x ⫽ [1, 0]T ⫹ ␭[1, 1]T and yields the point (␭* ⫽ ⫺–56 ); y (2) ⫽ [–16 , ⫺–56 ]T. According to the parallel subspace property, y (2) ⫺ y (1) ⫽ [–61 , ⫺–65 ]T ⫺ [⫺–61 , ⫺–61 ]T ⫽ [–31 , ⫺–32 ]T is conjugate to d ⫽ [1, 1]T. By definition, this means that [1, 1]C[–13 , ⫺–23 ]T ⫽ 0; which it is. Note that according to our previous observation, since we are dealing with a two-variable function, if a single-variable search is now performed along the line (y (2) ⫺ y (1)) the optimum of q(x) will be found. This is readily verified, since the minimum along the line x ⫽ [⫺–16 , ⫺–16 ]T ⫹ ␭[–13 , ⫺–23 ]T 1 3 1 T is (␭* ⫽ ⫺–– 16 ); x* ⫽ [⫺–– 16 , ⫺– 8 ] , which is the solution previously obtained.

In the previous construction, we require the initial specification of two points and a direction in order to produce a single conjugate direction. This is clumsy from a computational point of view, and we would prefer to generate the conjugate directions from a single starting point. This is easily accomplished by employing the coordinate unit vectors e(1), e(2), e(3), . . . , e(N) as follows. (Here we demonstrate the procedure in two dimensions; the extension to N dimensions should be obvious.) Let e(1) ⫽ [1, 0]T and e(2) ⫽ [0, 1]T. Given a starting point x (0), calculate ␭(0) such that ƒ(x(0) ⫹ ␭(0)e(1)) is minimized. Let x (1) ⫽ x (0) ⫹ ␭(0)e(1) and calculate ␭(1) so that ƒ(x (1) ⫹ ␭(1)e(2)) is minimized, and let

104


x (2) ⫽ x (1) ⫹ ␭(1)e(2) Next calculate ␭(2) so that ƒ(x (2) ⫹ ␭(2)e(1)) is minimized, and let x (3) ⫽ x (2) ⫹ ␭(2)e(1) Then, the directions x (3) –x (1) and e(1) will be conjugate. To see this, consider Figure 3.11. Note that the point x (1) is found by searching along e(1) from x (0), and x (3) is found by searching along e(1) from x (2). Consequently, by the parallel subspace property, the directions e(1) and x (3) –x (1) are conjugate. Furthermore, if we continue the iteration with a search along the line x (3) –x (1), we will have searched along two conjugate directions, and since we assume that ƒ(x) is a two-dimensional quadratic, the solution x* will be found. The preceding construction and the conjugacy property upon which it is based were stated in terms of only two conjugate directions. However, both can be readily extended to more than two conjugate directions and hence to higher dimensions. In particular, it is easy to show that if from x (1) the point y (1) is found after searches along each of M(⬍N) conjugate directions and, similarly, if from x (2) the point y (2) is found after searches along the same M conjugate directions, s (1), s (2), s (3), . . . , s (M), then the vector y (2) –y (1) will be conjugate to all of the M previous directions. This is known as the extended parallel subspace property. Using this extension, the construction shown in Figure 3.11 can immediately be generalized to higher dimensions. A threedimensional case is shown in Figure 3.12.

Figure 3.11. Conjugacy from a single point.

3.2


105

Figure 3.12. Construction of conjugate directions in three dimensions.

In the construction shown, the search begins with the three coordinate directions e(1), e(2), and e(3), and these are replaced in turn by newly generated conjugate directions. From x (0) a cycle of line searches is made along e(3), e(1), e(2), and again e(3). At the conclusion of this cycle the directions e(3) and x (4) –x (1) will be conjugate. The new search direction designated ④ in Figure 3.12 now replaces e(1). A new cycle of line searches is executed using directions ④, e(2), e(3), and again ④. Once again, by the extended parallel subspace property, the new direction x (8) –x (5), designated ⑤ in the figure, will be conjugate not only to ④ but also to e(3). Hence, the set of directions e(3), x (4) – x (1), and x (8) –x (5) are mutually conjugate. Therefore, if one additional line search is executed from x (8) along x (8) –x (5) —that is, ⑤ in the figure—the point x (9) is found, which must be the optimum if ƒ(x) is a three-dimensional quadratic, since we have searched three mutually conjugate directions in turn. Thus we see that in three dimensions nine line searches using only function values

106


are required to determine exactly (assuming exact arithmetic, of course) the optimum of a quadratic function. The construction is easily generalized and will in N dimensions require N 2 line searches to optimize a quadratic function. The generalization to N dimensions is outlined below. Powell’s Conjugate Direction Method

Step 1. Define x (0), the starting point, and a set of N linearly independent directions; possibly s (i) ⫽ e(i), i ⫽ 1, 2, 3, . . . , N. Step 2. Minimize along the N ⫹ 1 directions using the previous minimum to begin the next search and letting s (N ) be the first and last searched. Step 3. Form the new conjugate direction using the extended parallel subspace property. Step 4. Delete s (1) and replace it with s (2), and so on. Place the new conjugate direction in s (N ). Go to step 2. For the method (as described here) to be of practical use, a convergence check must be inserted as well as a test to ensure linear independence of the direction set. The independence check is particularly important when ƒ(x) is not quadratic [18, 19]. If follows from the way the algorithm is constructed that if the function is quadratic and has a minimum, then this minimum will be reached in exactly N loops through steps 2–4, where N is the number of variables. If the function is not quadratic, then of course more than N loops will be required. However, it can be proven rigorously under a reasonable hypothesis that the method will converge to a local minimum and will do this at a superlinear rate (as defined below). Rate of Convergence. The methods considered in this text generate a sequence of iterates x (k) that approximate the solution x*. We say a method is convergent if 㛳␧ (k⫹1)㛳 ⭐1 㛳␧ (k)㛳

(3.39)

␧ (k) ⫽ x (k) ⫺ x*

(3.40)

holds at each iteration, where

Since we are performing calculations using finite arithmetic, even the best algorithms are doomed to generate an infinite sequence of iterates. Accordingly, we are interested in the asymptotic convergence properties of these methods. Therefore, by convention we say (see Gill et al. [20] or Fletcher [21] for more complete discussions) that an algorithm has r order (or rate) of convergence if

3.2

lim k→⬁


㛳␧ (k⫹1)㛳 ⫽C 㛳␧ (k)㛳r

107

(3.41)

where C is the error constant. From the above we see that C ⭐ 1 when r ⫽ 1. When r ⫽ 1 we say the algorithm exhibits a linear convergence rate, and when r ⫽ 2 we say the rate is quadratic. When r ⫽ 1 and C ⫽ 0, the algorithm has a superlinear rate of convergence. Example 3.6

Powell’s Conjugate Direction Method

Find the minimum of ƒ(x) ⫽ 2x 31 ⫹ 4x1x 32 ⫺ 10x1x2 ⫹ x 22 from the starting point x (0) ⫽ [5, 2]T, where ƒ(x (0)) ⫽ 314. Step 1. s (1) ⫽ [1, 0]T and s (2) ⫽ [0, 1]T. Step 2. (a) Find ␭ such that ƒ(x (0) ⫹ ␭s (2)) → min. The result is ␭* ⫽ ⫺0.81, so x (1) ⫽ [5, 2]T ⫺ 0.81[0, 1]T ⫽ [5, 1.19]T

ƒ(x (1)) ⫽ 250

(b) Find ␭ such that ƒ(x (1) ⫹ ␭s (1)) → min. ␭* ⫽ ⫺3.26

x (2) ⫽ [1.74, 1.19]T

ƒ(x (2)) ⫽ 1.10

(c) Find ␭ such that ƒ(x (2) ⫹ ␭s (2)) → min. ␭* ⫽ ⫺0.098

x (3) ⫽ [1.74, 1.092]

ƒ(x (3)) ⫽ 0.72

Step 3. Let s (3) ⫽ x (3) ⫺ x (1) ⫽ [⫺3.26, ⫺0.098]T. Normalize. s (3) ⫽

s (3) ⫽ [⫺0.99955, ⫺0.03]T 㛳s (3)㛳

Let s (1) ⫽ s (2), s (2) ⫽ s (3), and go to step 2. Step 4. Find ␭ such that ƒ(x (3) ⫹␭s (2)) → min. ␭* ⫽ 0.734

x (4) ⫽ [1.006, 1.070]T

ƒ(x (4)) ⫽ ⫺2.86

Note: If ƒ(x) were a quadratic, the problem would be solved at this point (ignoring round-off error); since it is not, the iteration continues to the solution. The search directions and iterates are shown in Figure 3.13.

108


Figure 3.13. Powell’s conjugate directions method, Example 3.6.

All available computational evidence [22] indicates that Powell’s method (with the linear dependence check added) is at least as reliable as any other direct-search method and is usually much more efficient. Consequently, if a sophisticated algorithm is in order, this method is clearly to be recommended. This completes our discussion of direct-search methods for unconstrained functions. In the next section we consider methods that employ derivative information.

3.3

GRADIENT-BASED METHODS

In the previous section we examined several methods that require only objective function values to proceed toward the solution. These direct methods are important, because very often in a practical engineering problem this is the only reliable information available. On the other hand, even the best direct methods can require an excessive number of function evaluations to locate the solution. This, combined with the quite natural desire to seek stationary

3.3


109

points [i.e., points that satisfy the first necessary condition, (3.15a)], motivates us to consider methods that employ gradient information. The methods will be iterative, since the elements of the gradient will in general be nonlinear functions of the design variables. We shall assume throughout that ƒ(x), ⵜƒ(x), and ⵜ2ƒ(x) exist and are continuous. We briefly consider methods that require both first and second derivatives, but primarily as a point of reference and as motivation for the more useful methods. In particular, we will examine in detail conjugate gradient methods, which exploit the conjugacy concept previously discussed but by use of gradient information, and quasi-Newton methods, which mimic Newton’s method but once again require only first-order information. We assume that the elements of the gradient are available in closed form or can be reliably approximated numerically. In addition, we give suggestions for making numerical approximations to the gradient. All of the methods considered here employ a similar iteration procedure: x (k⫹1) ⫽ x (k) ⫹ ␣ (k)s(x (k))

(3.42)

where x (k) ⫽ current estimate of x*, the solution ␣ (k) ⫽ step-length parameter s(x (k)) ⫽ s (k) ⫽ search direction in the N space of the design variables xi, i ⫽ 1, 2, 3, . . . , N The manner in which s(x) and ␣ are determined at each iteration defines a particular method. Usually ␣ (k) is selected so as to minimize ƒ(x) in the s(x (k)) direction. Therefore, we will require efficient single-variable minimization algorithms to implement these methods. 3.3.1

Cauchy’s Method [23]

Let us assume for the moment that we are stationed at the point x in the design space and wish to determine the direction of ‘‘most local descent.’’ As before, consider the Taylor expansion of the objective about x: ƒ(x) ⫽ ƒ(x) ⫹ ⵜƒ(x)T ⌬ x ⫹ 䡠䡠䡠

(3.43)

where we ignore terms of degree 2 and higher. Clearly the second term will dictate the local decrease in the objective value, since ƒ(x) is fixed. The greatest decrease in ƒ is associated with the choice of direction in (3.42) that produces the greatest negative scalar product in the second term. The greatest negative scalar product results from the choice s(x) ⫽ ⫺ⵜƒ(x)

(3.44)

clearly from the projection property of scalar products. The second term then becomes

110


⫺␣ ⵜƒ(x)T ⵜƒ(x)

producing the most local descent. This is the motivation for the simple gradient method: x (k⫹1) ⫽ x (k) ⫺ ␣ ⵜƒ(x (k))

(3.45)

where ␣ is a fixed positive parameter. The method has two disadvantages: the need to make an appropriate choice for ␣ and inherent sluggishness near the minimum due to the corrections vanishing as ⵜƒ goes to zero. Accordingly, we are led to adjust ␣ at each iteration: x (k⫹1) ⫽ x (k) ⫺ ␣ (k) ⵜƒ(x (k))

(3.46)

We determine ␣ (k) such that ƒ(x (k⫹1)) is a minimum along ⵜƒ(x (k)) using an appropriate single-variable searching method. We call this the method of steepest descent or Cauchy’s method, since Cauchy first employed a similar technique to solve systems of linear equations. The line search implied in (3.46) will make the method more reliable than the simple gradient method, but still the rate of convergence will be intolerably slow for most practical problems. It seems reasonable that the method will approach the minimum slowly, since the change in the variables is directly related to the magnitude of the gradient, which is itself going to zero. There is, then, no mechanism to produce an acceleration toward the minimum in the final iterations. One major advantage of Cauchy’s method is its surefootedness. The method possesses a descent property, since by the very nature of the iteration we are assured that ƒ(x (k⫹1)) ⭐ ƒ(x (k))

(3.47)

With this in mind we note that the method typically produces good reduction in the objective from points far from the minimum. It is then a logical starting procedure for all gradient-based methods. Finally, the method does demonstrate many of the essential ingredients of all methods in this class. Example 3.7

Cauchy’s Method

Consider the function ƒ(x) ⫽ 8x 21 ⫹ 4x1x2 ⫹ 5x 22 to which we wish to apply Cauchy’s method.

3.3


111

Solution. To begin, calculate the elements of the gradient: ⭸ƒ ⫽ 16x1 ⫹ 4x2 ⭸x1

⭸ƒ ⫽ 10x2 ⫹ 4x1 ⭸x2

Now to use the method of steepest descent, we estimate the solution, x (0) ⫽ [10, 10]T and generate a new estimate using expression (3.46): x (1) ⫽ x (0) ⫺ ␣ (0) ⵜƒ(x (0)) Now we choose ␣ (0) such that ƒ(x (1)) → min; ␣ (0) ⫽ 0.056. Therefore, x (1) ⫽ [⫺1.20, 2.16]T. Next we calculate x (2) ⫽ x (1) ⫺ ␣ (1) ⵜƒ(x (1)) by evaluating the gradient at x (1) and performing the line search. If we perform the line search using the quadratic interpolation method [24, 25], we get the iterates given in Table 3.1 and displayed in Figure 3.14. Cauchy’s method, even though not of great practical value, exhibits the major steps of most gradient-based methods. A flowchart of the algorithm is given in Figure 3.15. Note that termination tests include the length of the gradient and ⌬ x vectors. 3.3.2

Newton’s Method [26]

We have seen that Cauchy’s method is the ‘‘best’’ local gradient-based strategy, but observe that the negative gradient points directly toward the minimum only when the contours of ƒ are circular, and therefore the negative gradient is not a good global direction (in general) for nonlinear functions. Cauchy’s method employs successive linear approximations to the objective and requires values of the objective and gradient at each iteration. We are then led

Table 3.1

k 0 1 2 3 4

Cauchy Method

x (k) 1 10.0 ⫺1.2403

0.1441 ⫺0.0181

0.0021

x (k) 2

ƒ(x (k))

10.0 2.1181 0.1447 0.0309 0.0021

1700 24.2300 0.3540 0.0052 0.0000

112


Figure 3.14. Cauchy / quadratic interpolation iterates.

to consider the use of higher order information, namely second derivatives, in an effort to construct a more global strategy. Consider once again the Taylor expansion of the objective: ƒ(x) ⫽ ƒ(x (k)) ⫹ ⵜƒ(x (k))T ⵜx ⫹ –21 ⌬ xT ⵜ2ƒ(x (k)) ⌬ x ⫹ O(⌬ x 3) We form a quadratic approximation to ƒ(x) by dropping terms of order 3 and above: ˜ ƒ(x; x (k)) ⫽ ƒ(x (k)) ⫹ⵜƒ(x (k))T ⌬ x ⫹ –21 ⌬ xT ⵜ2ƒ(x (k)) ⌬ x

(3.48)

˜ where we use ƒ(x; x (k)) to denote an approximating function constructed at (k) x , which is itself a function of x. Now let us use this quadratic approximation of ƒ(x) to form an iteration sequence by forcing x (k⫹1), the next point in the sequence, to be a point where the gradient of the approximation is zero. Therefore,

3.3


113

Figure 3.15. Cauchy’s method.

˜ ⵜƒ(x; x (k)) ⫽ ⵜƒ(x (k)) ⫹ ⵜ2ƒ(x (k)) ⌬ x ⫽ 0

(3.49)

⌬ x ⫽ ⫺ⵜ2ƒ(x (k))⫺1 ⵜƒ(x (k))

(3.50)

so

Accordingly, this successive quadratic approximation scheme produces Newton’s optimization method:

114


x (k⫹1) ⫽ x (k) ⫺ ⵜ2ƒ(x (k))⫺1 ⵜƒ(x (k))

(3.51)

which we demonstrate in the following example. Example 3.8

Newton’s Method

Consider once again the function given in the previous example: ƒ(x) ⫽ 8x 21 ⫹ 4x1x2 ⫹ 5x 22 where ⵜƒ(x) ⫽ [16x1 ⫹ 4x2, 4x1 ⫹ 10x2]T

and ⵜ2ƒ(x) ⫽ H ⫽

冋

16 4 4 10

册

Using (3.51), with x (0) ⫽ [10, 10]T as before, we get x (1) ⫽ [10, 10]T ⫺

冉冊冋

x (1) ⫽ [10, 10]T ⫺

冉冊

1 144

册

10 ⫺4 [200, 140]T 16

⫺4

and therefore 1 [1440, 1440]T ⫽ [0, 0]T 144

which is the exact solution. We see, as we might expect from the previous example, that Newton’s method will minimize a quadratic function (from any starting point) in exactly one step. Convergence Properties. It is worthwhile to consider the convergence prop-

erties of Newton’s method. Mangasarian [27] has shown, under rather mild regularity conditions on ƒ(x), that Newton’s method exhibits a quadratic rate of convergence: 㛳␧ (k⫹1)㛳 ⭐ C㛳␧ (k)㛳2

(3.52)

where the error constant C ⫽ (R/2M), R being the upper bound on the L2 norm of ⵜ2ƒ and M the condition number of ⵜ2ƒ⫺1. Clearly, Newton’s method will converge whenever x (0) is chosen such that

3.3

㛳␧(x (0))㛳 ⬍


1 C

115

(3.53)

We might have expected the quadratic convergence rate, since the method is based on a quadratic approximation. For arbitrary functions it is entirely feasible to assume that we might choose a point such that 㛳␧(x (0))㛳 ⬎ 1/C, and the method may perform poorly. Note that the algorithm does not possess a descent property. To see this more clearly, assume the current iterate x is not a stationary point [i.e., ⵜƒ(x) ⫽ 0], and examine the projection of the Newton direction onto the local gradient. By definition, the search direction is a descent direction if ⵜƒ(x)Ts(x) ⬍ 0

(3.54)

Therefore, from (3.51) this leads to ⫺ⵜƒ(x)T ⵜ2ƒ(x)⫺1 ⵜƒ(x) ⬍ 0

(3.55)

Clearly this condition is satisfied, and the Newton direction is ‘‘downhill’’ if ⵜ2ƒ(x) is positive definite. If we encounter a point where ⵜ2ƒ(x) is negative definite, then the direction points ‘‘uphill’’; and when the Hessian is indefinite, there is no assurance of ascent or descent. 3.3.3

Modified Newton’s Method

Experience has shown that Newton’s method can be unreliable for nonquadratic functions. That is, the Newton step will often be large when x (0) is far from x*, and there is the real possibility of divergence. It is possible to modify the method in a logical and simple way to ensure descent by adding a line search as in Cauchy’s method. That is, we form the sequence of iterates x (k⫹1) ⫽ x (k) ⫺ ␣ (k) ⵜ2ƒ(x (k))⫺1 ⵜƒ(x (k))

(3.56)

by choosing ␣ (k) such that ƒ(x (k⫹1)) → min which ensures that ƒ(x (k⫹1)) ⭐ ƒ(x (k)) This is the modified Newton’s method, and we find it reliable and efficient when first and second derivatives are accurately and inexpensively calculated. The major difficulty is the need to calculate and solve the linear equation set involving the Hessian ⵜ2ƒ(x) at each iteration.

116


3.3.4

Marquardt’s Method

This method combines Cauchy’s and Newton’s methods in a convenient manner that exploits the strengths of both but does require second-order information. We have previously noted that the gradient is the direction of most local ascent and Cauchy’s method typically provides a good reduction in the objective when x (0) is far from x(*). On the other hand, Newton’s method generates ideal search directions near the solution. It is this simple idea that motivates the method that Marquardt introduced in 1963 [28]. He specified the search direction to be s(x (k)) ⫽ ⫺[H(k) ⫹ ␭(k)I]⫺1 ⵜƒ(x (k))

(3.57)

and set ␣ (k) ⫽ ⫹1 in (3.42), since ␭ is used to control both the direction of search and the length of the step. Here I is the identity matrix, that is, it contains zeros except for the diagonal elements, which are ⫹1. To begin the search, let ␭(0) be a large constant, say 104, such that [H(0) ⫹ ␭(0)I]⫺1 艑 [␭(0)I]⫺1 ⫽

冉冊

1 I ␭(0)

(3.58)

Therefore, for ␭(0) sufficiently large, s(x (0)) → ⫺ⵜƒ(x (0)). We notice from (3.57) that as ␭ decreases from a large value to zero, s(x) goes from the gradient to the Newton direction. Accordingly, we test descent after the first step, and when the step is a success [that is, ƒ(x (1)) ⬍ ƒ(x (0))], we let ␭(1) ⬍ ␭(0) and step again; otherwise we set ␭(0) ⫽ ␤␭(0), where ␤ ⬎ 1, and perform the previous step again. The algorithm is summarized below. Marquardt’s Compromise

Step 1. Define x (0) ⫽ initial estimate of x* M ⫽ maximum number of iterations allowed ␧ ⫽ convergence criterion Step 2. Set k ⫽ 0. ␭(0) ⫽ 104. Step 3. Calculate ⵜƒ(x (k)). Step 4. Is 㛳ⵜƒ(x (k))㛳 ⬍ ␧? Yes: Go to step 11. No: Continue. Step 5. Is k ⭓ M? Yes: Go to step 11. No: Continue. Step 6. Calculate s(x (k)) ⫽ ⫺[H(k) ⫹ ␭(k)I]⫺1 ⵜƒ(x (k)). Step 7. Set x (k⫹1) ⫽ x (k) ⫹ s(x (k)).

3.3


117

Step 8. Is ƒ(x (k⫹1)) ⬍ ƒ(x (k))? Yes: Go to step 9. No: Go to step 10. Step 9. Set ␭(k⫹1) ⫽ –12 ␭(k) and k ⫽ k ⫹ 1. Go to step 3. Step 10. Set ␭(k) ⫽ 2␭(k). Go to step 6. Step 11. Print results and stop. The major advantages of the method are its simplicity, the descent property, the excellent convergence rate near x*, and the absence of a line search. The major disadvantage is the need to calculate H(k) and then solve the linear equation set corresponding to (3.57). This method has been used quite extensively with problems where ƒ(x) is a sum of squares, that is, ƒ(x) ⫽ ƒ21(x) ⫹ ƒ22(x) ⫹ 䡠䡠䡠 ⫹ ƒm2 (x)

(3.59)

(In fact this is the problem originally considered by Marquardt.) Powell [29] and Bard [30] give numerical results that indicate that the method is particularly attractive for this application. 3.3.5

Conjugate Gradient Methods

In the previous sections we have discussed Cauchy’s and Newton’s methods. We have seen that Cauchy’s method tends to be effective far from the minimum but becomes less so as x* is approached, whereas Newton’s method can be unreliable far from x* but is very efficient as x (k) approaches the minimum. In this and the following sections we discuss methods that tend to exhibit the positive characteristics of the Cauchy and Newton methods using only firstorder information. That is, the methods to be considered here will be reliable far from x* and will accelerate as the sequence of iterates approaches the minimum. We call these methods quadratically convergent if they terminate in approximately N steps when applied to a quadratic function and exact arithmetic is employed. One such class of methods comprises those that employ conjugate directions. We have previously given the conjugacy conditions for a set of directions s (k), k ⫽ 1, 2, 3, . . . , r ⭐ N, and an N ⫻ N symmetric matrix C. In addition, we have examined these directions as transformations into the form of a sum of perfect squares for an arbitrary quadratic function. In this way we concluded that searching in each of N C-conjugate directions in turn would produce the minimum for an N-dimensional quadratic. Furthermore, we generated the set of conjugate directions using values of the objective only. Here a quadratic approximation to ƒ(x) and gradient information are employed to generate conjugate directions. In addition, we shall insist upon descent at each iteration.

118


Consider any two distinct points in the design space x (0) and x (1) and the gradient of the quadratic objective at these points: ⵜƒ(x) ⫽ ⵜq(x) ⫽ Cx ⫹ b ⫽ g(x)

(3.60)

where for convenience we use the symbol g(x) for the gradient. Therefore, g(x (0)) ⫽ Cx (0) ⫹ b g(x (1)) ⫽ Cx (1) ⫹ b Now form the change in the gradient from x (0) to x (1): ⌬g(x) ⫽ g(x (1)) ⫺ g(x (0)) ⫽ C(x (1) ⫺ x (0)) ⫽ C ⌬x

(3.61)

which is the property of quadratic functions we shall exploit. In 1952, Hestenes and Stiefel [31] published an efficient iterative technique for solving systems of linear equations, which is essentially the method of conjugate gradients. They viewed the set of linear equations as elements of the gradient vector of a quadratic to which they sought the minimum. Later, Fletcher and Reeves [32] proved the quadratic convergence and extended the method to nonquadratic functions. Fried and Metzler [33] have demonstrated the utility of the method when it is applied to linear sets resulting from finiteelement discretizations, where the coefficient matrix is sparse. They stress the ease of implementing the method as compared to other more commonly used techniques. This is an attractive feature to us as well. We assume in the development of the method that the objective is quadratic, ƒ(x) ⫽ q(x) ⫽ a ⫹ bTx ⫹ –12 xTCx and the iteration prescription in (3.42) is employed, namely, x (k⫹1) ⫽ x (k) ⫹ ␣ (k)s(x (k)) Let us form the search directions at each iteration using

冘␥

k⫺1

s (k) ⫽ ⫺g (k) ⫹

i⫽0

with

(i) (i)

s

(3.62)

3.3


s (0) ⫽ ⫺g (0)

119

(3.63)

where g (k) ⫽ ⵜƒ(x (k)). Since we shall perform line searches along each direction in turn as they are formed, it is useful to recall that the following condition holds as a result of termination of the line search: ⵜƒ(x (k⫹1))Ts (k) ⫽ 0

(3.64)

Now we wish to choose ␥ (i), i ⫽ 1, 2, 3, . . . , k ⫺ 1, such that s (k) is C-conjugate to all previous search directions. Consider the first direction: s (1) ⫽ ⫺g (1) ⫹ ␥ (0)s (0) ⫽ ⫺g (1) ⫺ ␥ (0)g (0) and force conjugacy with s (0) only: s (1)TCs (0) ⫽ 0 [g (1) ⫹ ␥ (0)g (0)]TCs (0) ⫽ 0 From the previous iteration, we know that s (0) ⫽

⌬x ␣ (0)

and therefore

冋册

[g (1) ⫹ ␥ (0)g (0)]TC

⌬x ⫽0 ␣ (0)

From the quadratic property (3.61) we get [g (1) ⫹ ␥ (0)g (0)]T ⌬g ⫽ 0

(3.65)

or ␥ (0) ⫽

⫺⌬gT g (1) ⌬gT g (0)

Expansion of (3.65) produces g (1)Tg (1) ⫹ ␥ (0)g (0)Tg (1) ⫺ g (1)Tg (0) ⫺ ␥ (0)g (0)Tg (0) ⫽ 0 but by choice of ␣ (0) and (3.64) we know that

(3.66)

120


g (1)Tg (0) ⫽ 0 so that ␥ (0) ⫽

㛳 g (1)㛳2 㛳 g (0)㛳2

(3.67)

Continuing the process, we form the next search direction: s (2) ⫽ ⫺g (2) ⫹ ␥ (0)s (0) ⫹ ␥ (1)s (1) and choose ␥ (0) and ␥ (1) such that s (2)TCs (0) ⫽ 0

and

s (2)TCs (1) ⫽ 0

s (2) is C-conjugate to s (0) and s (1). We leave it as an exercise to show that ␥ (0) ⫽ 0 here and ␥ (i) ⫽ 0 for i ⫽ 0, 1, 2, . . . , k ⫺ 2 for all k in general using the previous reasoning and conditions (3.61) and (3.64). This leads to the general iteration s (k) ⫽ ⫺g (k) ⫹

冋

册

㛳 g (k)㛳2 s (k⫺1) 㛳 g (k⫺1)㛳2

(3.68)

which is the form suggested by Fletcher and Reeves. When ƒ(x) is a quadratic, N ⫺ 1 such directions and N line searches will be required (in the absence of round-off error) to locate the minimum. When ƒ(x) is not quadratic, additional directions and line searches are required. Several investigators and some computational experience suggest it is wise to restart [i.e., set s(x) ⫽ ⫺g(x)] every N or N ⫹ 1 steps. Regular restarts do seem to provide a safer procedure, since the generation of a linearly dependent direction is less likely. The generation of a direction that is linearly dependent upon one or more of the previous directions may cause the method to fail, since a subspace of RN is searched thereafter. Quite happily, this seldom occurs in practice. In fact the method has performed quite well in practice and is especially appealing because of the simplicity of the one-term update required and the small amount of storage needed to advance the search. As the number of variables grows large, the method and its several variants become increasingly attractive due to the small relative storage requirement. Example 3.9

Fletcher–Reeves Method

Find the minimum of the function ƒ(x) ⫽ 4x 21 ⫹ 3x 22 ⫺ 4x1x2 ⫹ x1 from x (0) ⫽ [0, 0]T.

3.3


121

Solution Step 1. ⵜƒ(x) ⫽ [8x1 ⫺ 4x2 ⫹ 1 s (0) ⫽ ⫺ⵜƒ(x (0)) ⫽ ⫺[1, 0]T Step 2. Line search:

6x2 ⫺ 4x1]T

x (1) ⫽ x (0) ⫺ ␣ (0) ⵜƒ(x (0)) → ␣ (0) ⫽ –81 x (1) ⫽ [0, 0]T ⫺ –18 [1, 0] ⫽ [⫺–18 , 0]T Step 3. k ⫽ 1: s (1) ⫽ ⫺[0, –12 ]T ⫺ (–14 )[1, 0]T ⫽ ⫺[–14 , –12 ]T Step 4. Line search: x (2) ⫽ x (1) ⫹ ␣ (1)s (1) → ␣ (1) ⫽ –41 3 1 T x (2) ⫽ [⫺–81 , 0]T ⫺ –41 [–41 , –21 ]T ⫽ [⫺–– 16 , ⫺– 8]

ⵜƒ(x (2)) ⫽ [0, 0]T

Therefore, x (2) ⫽ x*, and the solution is achieved in two line searches, because they were performed exactly and the objective is quadratic. Miele and Cantrell [34] have given an extension of the Fletcher–Reeves approach: x (k⫹1) ⫽ x (k) ⫹ ␣ (k){⫺ⵜƒ(x (k)) ⫹ ␥ (k)s(x (k⫺1))}

(3.69)

where ␣ (k) and ␥ (k) are sought directly at each iteration. The method, known as the memory gradient method, is remarkably efficient in terms of number of iterations but requires more function and gradient evaluations than Fletcher–Reeves (FR). Accordingly, the method seems useful only when objective and gradient evaluations are very inexpensive. Recall that in the development of the FR method both a quadratic objective and exact line searches were assumed. A host of methods have been proposed that generate conjugate directions while relaxing one or both of these conditions. Possibly the most noteworthy is given by Polak and Ribiere [35]. This method demands exact line searches but assumes a more general objective model, to give ␥ (k) ⫽

where

⌬g(x (k))Tg(x (k)) 㛳 g(x (k⫺1)㛳2

(3.70)

122


⌬g(x (k)) ⫽ g(x (k)) ⫺ g(x (k⫺1))

(3.71)

as before. Using ␣ supplied by the line search and ␥ in (3.70), we have the Polak and Ribiere conjugate gradient method: x (k⫹1) ⫽ x (k) ⫹ ␣ (k)s(x (k)) s(x (k)) ⫽ ⫺ⵜƒ(x (k)) ⫹ ␥ (k)s(x (k⫺1))

(3.72)

Quite obviously, the only difference between the Polak–Ribiere and FR methods is the choice of ␥. Other similar methods have been proposed that assume exact line searches but employ a more general (than quadratic) objective model. See, for instance, Davison and Wong [36] and Boland et al. [37, 38]. In 1972, Crowder and Wolfe [39] and later Powell [40] proved that conjugate gradient methods were doomed to a linear rate of convergence in the absence of periodic restarts. A restart is defined as a special procedure that interrupts the normal generation of search directions, like that which is required for s(x (0)). As we have said, there are several reasons one might expect an algorithm to require a restart procedure in order to be robust. The most obvious is to avoid the generation of a dependent direction. Powell [41] proved that the Polak–Ribiere method also gives a linear rate of convergence without restart but is superior to the FR method for general functions and exhibits less sensitivity to inexact line searches. The formulation of efficient restart procedures and methods with greater tolerance of inexact line searches remain active areas of research. Beale [42] has given a conjugate gradient procedure similar to the standard FR method that allows restarting without the use of the gradient direction. He shows how the direction just prior to restart might be used to reduce the overall computation effort for functions that require many restarts. Powell [41] carefully examines Beale’s and other restarting strategies and suggests that a restart should occur every N steps, or when 兩g(x (k))g(x (k⫺1))兩 ⭓ 0.2㛳 g(x (k))㛳2

(3.73)

He demonstrates that Beale’s restart strategy [using his criterion (3.73)] works equally well with the FR or Polak–Ribiere direction formulas and gives numerical results that suggest the superiority of the Polak–Ribiere (with restart) method. Shanno [43] has studied the relation of inexact line searches and restart strategies on the efficiency of conjugate gradient methods. He shows that Beale’s restart strategy (along with the associated two-term update formula) with Powell’s restart criterion allows a significant decrease in the required precision of the line searches and therefore a significant increase in the overall computational efficiency of the conjugate gradient method. Shanno gives numerical results that indicate the superiority of the Polak–Ribiere

3.3


123

method with restart and inexact line searches. Gabriele and Ragsdell [44] demonstrated the value of these techniques for large nonlinear programs, where their primary value remains. 3.3.6

Quasi-Newton Methods

These methods are similar to those of Section 3.3.5 in that they are based primarily upon properties of quadratic functions. Whereas the previous methods drew their strength from conjugate search directions, the current methods are designed to more directly mimic the positive characteristics of Newton’s method using only first-order information. All methods of this class generate directions to be used in (3.42) of the form s(x (k)) ⫽ ⫺A(k) ⵜƒ(x (k))

(3.74)

where A(k) is an N ⫻ N matrix called the metric. Methods that employ directions of this form are often called variable metric methods, because A changes at each iteration. To be precise, a variable metric method is a quasi-Newton method if it is designed so that the iterates satisfy the following quadratic property: ⌬ x ⫽ C⫺1 ⌬g

(3.75)

Unfortunately, the literature is not precise or consistent in the use of these terms [45]; therefore, we shall use them interchangeably. This is appropriate, since both expressions are of equal importance in the design and execution of these methods. Let us assume a recursion for the estimate to the inverse of the Hessian: A(k⫹1) ⫽ A(k) ⫹ A(k) c

(3.76)

(k) where A(k) in (3.74) and c is a correction to the current metric. We will use A (k) (0) (3.42). The plan is to form Ac such that the sequence A , A(1), A(2), . . . , A(k⫹1) approaches H⫺1 ⫽ ⵜ2ƒ(x*)⫺1, for in so doing one additional line search will produce x* if ƒ(x) is quadratic. As we have said several times before, we hope that success with quadratic functions will lead to success with general nonlinear functions, and experience suggests that this hope is often realized. Recall the important property of a quadratic function given in (3.75), and assume our approximation of C⫺1 is of the form

C⫺1 ⫽ ␤A(k)

(3.77)

where ␤ is a scalar. We would like the approximation to satisfy (3.75), that is,

124


⌬ x (k) ⫽ ␤A(k) ⌬g (k)

(3.78)

but this is clearly impossible, since we need A(k) to find ⌬g (k), where ⌬ x (k) ⫽ x (k⫹1) ⫺ x (k) ⌬g

(k)

⫽ g(x

(3.79)

) ⫺ g(x )

(k⫹1)

(k)

(3.80)

On the other hand, we can require that the new approximation satisfy (3.75): ⌬ x (k) ⫽ ␤A(k⫹1) ⌬g (k)

(3.81)

Combining (3.76) and (3.81) leads to (k) A(k) ⫽ c ⌬g

冉冊

1 ⌬ x (k) ⫺ A(k) ⌬g (k) ␤

(3.82)

A solution is not obvious, but one can verify by direct substitution that A(k) c ⫽

冉

冊

1 ⌬ x (k) yT A(k) ⌬g (k) zT ⫺ T (k) ␤ y ⌬g zT ⌬g (k)

(3.83)

is a solution. Here y and z are arbitrary vectors, so this is really a family of solutions. If we let y ⫽ ⌬ x (k)

and

z ⫽ A(k) ⌬g (k)

(3.84)

we get the well-known and widely used Davidon–Fletcher–Powell (DFP) method [46, 47]: A(k) ⫽ A(k⫺1) ⫹

⌬ x (k⫺1) ⌬ x(k⫺1)T A(k⫺1) ⌬g (k⫺1) ⌬g (k⫺1)T A(k⫺1) ⫺ ⌬ x(k⫺1)T ⌬g (k⫺1) ⌬g (k⫺1)T A(k⫺1) ⌬g (k⫺1)

(3.85)

It can be shown that this updating formula preserves symmetry and positive definiteness, so that if A(0) is any symmetric positive-definite matrix, A(1), A(2), . . . will also be symmetric and positive definite in the absence of roundoff error. It is convenient to let A(0) ⫽ I. Recalling the first variation of ƒ(x), we write ⌬ƒ(x) ⫽ ⵜƒ(x (k))T ⌬ x

(3.86)

Now using (3.42) and (3.74) with this expression, we get ⌬ƒ(x) ⫽ ⫺ⵜƒ(x (k))T ␣ (k)A(k) ⵜƒ(x (k))

Therefore,

(3.87)

3.3


⌬ƒ(x) ⫽ ⫺␣ (k) ⵜƒ(x (k))T A(k) ⵜƒ(x (k))

125

(3.88)

assures us that ƒ(x (k⫹1)) ⬍ ƒ(x (k)) for all ␣ (k) ⬎ 0 and A(k) positive definite. So the algorithm has a descent property. The DFP method has been and continues to be a very widely used gradient-based technique. The method tends to be robust; that is, it will typically do well on a wide variety of practical problems. The major disadvantage associated with methods in this class is the need to store the N ⫻ N matrix A. Example 3.10

Davidon–Fletcher–Powell Method

Find the minimum of ƒ(x) ⫽ 4x 21 ⫹ 3x 22 ⫺ 4x1x2 ⫹ x1 from x (0) ⫽ [0, 0]T using the DFP method. Solution 1. Let s (0) ⫽ ⫺ⵜƒ(x (0)) ⫽ ⫺[1, 0]T. 2. Line search, same as in Example 3.9: x (1) ⫽ [⫺–81 , 0]T 3. k ⫽ 1: A(1) ⫽ A(0) ⫹ Ac(0) A(0) ⫽

冋册

A(0) c ⫽

⌬ x (0) ⌬ x (0)T A(0) ⌬g (0) ⌬g (0)T A(0) ⫺ (0)T (0) ⌬x ⌬g ⌬g (0)T A(0) ⌬g (0)

1 0

0 ⫽I 1

⌬ x (0) ⫽ [⫺–18 , 0]T ⫺ [0, 0]T ⫽ [⫺–18 , 0]T ⌬g (0) ⫽ [0, –21 ]T ⫺ [1, 0]T ⫽ [⫺1, –21 ]T

Ac(0) ⫽

A(0) c ⫽ A(1) ⫽

[⫺–, 0] [⫺–, 0] ⫺ [⫺–, 0][⫺1, –12 ]T 1 8 1 8

T

1 8

冋册冋冋册 –18

0

冋册 1 0

冋册

[⫺1, –12 ]

册

–45 ⫺–25 0 ⫺ –51 0 ⫺–52

0.325 –25 –25

冋册

0 1 [⫺1, –12 ]T[⫺1, –12 ] 1 0

–45

s (1) ⫽ ⫺A(1) ⵜƒ(x (1)) ⫽ ⫺[–51 , –52 ]T

1 0 [⫺1, –12 ]T 0 1

0 1

126


4. Line search: 5 x (2) ⫽ x (1) ⫹ ␣ (1)s (1) → ␣ (1) ⫽ –– 16 5 1 2 T 3 1 T x (2) ⫽ [⫺–81 , 0]T ⫺ –– 16 [– 5, – 5 ] ⫽ [⫺–– 16 , ⫺– 8]

which is the solution, as before. In Examples 3.9 and 3.10, it is noteworthy that the FR and DFP methods generated exactly the same search directions, and they are in fact H-conjugate:

冋

s (1)THs (0) ⫽ [–14 , –12 ]

册

⫺4

8

⫺4

6

[1, 0]T ⫽ 0

We also notice that the metric A(k) does converge to H⫺1:

冋

册

⫺4

8 ⫺4

H⫽

6

→ H⫺1 ⫽

冋册 3 –– 16 –18

–18 –14

3 1 T 1 T 1 1 T ⌬ x ⫽ x (2) ⫺ x (1) ⫽ [⫺–– 16 , ⫺– 8 ] ⫺ [⫺– 8 , 0] ⫽ [⫺–– 16 , ⫺– 8]

⌬g ⫽ g (2) ⫺ g (1) ⫽ [0, 0]T ⫺ [0, –21 ]T ⫽ [0, ⫺–21 ]T

(1) c

A

[⫺––, ⫺–] [⫺––, ⫺–] ⫽ ⫺ 1 [⫺–– 16 , ⫺–][0, ⫺–] 1 16

1 T 8 1 8

冋

1 8 1 T 2

冋

0.325 0.4

册

冋册册冋册

[0, ⫺–12 ]

册冋册冋

冋

册

0.4 0.325 0.4 [0, ⫺–12 ]T[0, ⫺–12 ] 0.8 0.4 0.8

1 –– –1 0.4 0.2 ⫹ 161 18 ⫺ – – 0.8 0.4 8 4

0.325 0.4

A(2) ⫽

1 16

0.325 0.4

0.4 [0, ⫺–12 ]T 0.8

3 –– –1 0.4 ⫽ 161 18 – – 0.8 8 4

So, A(2) ⫽ H⫺1(x*). Since the publication of the Davidon updating formula, there have been various other variable metric methods proposed that result from different choices for ␤, ␥, and z in (3.83). A common practical difficulty is the tendency of A(k⫹1) to be ill-conditioned (see Appendix C.3 for definition), causing increased dependence on a restarting procedure. Another method proposed almost simultaneously by Broyden [48], Fletcher [49], Goldfarb [50], and Shanno [51] has received wide acclaim and has in fact been strongly recommended by Powell [52]. The updating formula for the Broyden–Fletcher–Goldfarb–Shanno (BFGS) method is (k⫹1)

A

⫽

冋

册冋

册

⌬ x (k) ⌬g(k)T (k) ⌬ x (k) ⌬g (k)T I ⫺ (k)T A I⫺ (k) ⌬x ⌬g ⌬ x (k)T ⌬g (k)

T

⫹

⌬ x (k) ⌬ x (k)T ⌬ x (k)T ⌬g (k)

(3.89)

3.3


127

The major advantage of the BFGS method is the drecreased need to restart, and the method seems significantly less dependent upon exact line searching. There is a definite similarity between current efforts to improve variable metric methods and those to imporve conjugate gradient methods. Therefore, our remarks on conjugate gradient methods hold to some degree for variable metric methods as well. Davidon [53] has proposed a class of updates that permit inexact line searches. Powell [54] later established the quadratic termination property of these methods in the absence of line searches. It would appear that a variable metric method with quadratic termination but without the need of expensive line searches would be both robust and fast on general functions. Goldfarb [50] has also investigated this very promising possibility. In general, we restart the DFP method (that is, set A ⫽ I) after N updates. This is usually a very conservative measure, since many more steps may often be taken before a restart is required. We sense the need to restart as the condition number of the update worsens. Be aware that even though restarting does provide a degree of safety (or robustness) with variable metric methods, it does typically slow progress to the solution, since a second-order estimate is being discarded often the trade-off is between reliability and speed. McCormick [55], Shanno [43], and Shanno and Phua [56] have extensively investigated the relationship between conjugate gradient and variable metric methods. Shanno views conjugate gradient methods as ‘‘memoryless’’ variable metric methods. Most would now agree that these two classes of methods share much more than was originally thought. In addition, it is clear that the many variable metric variants have very much in common in actual practice [57, 58], which causes one to weigh carefully the additional computational overhead of the more complex methods. Shanno and Phua [59, 60] give extensive numerical results that favor the relatively simple BFGS update. Excellent software implementations of the best methods are available in the open literature [56, 61]. In addition, variable metric methods have been used extensively in the development of techniques for constrained problems. LaFrance et al. [62] have given a combined generalized reduced gradient and DFP method, whereas Root [63] used DFP with his implementation of the method of multipliers. Of course, the work of Han [64], Powell [65], and others [66] on constrained problems exploits the efficiency of variable metric methods. When N is large, we are usually forced to use some variant of the conjugate gradient method discussed in this chapter. In addition, we might consider using a variable metric method, which is designed to exploit sparsity in H or H⫺1 [67–70]. 3.3.7

Trust Regions

All gradient-based methods discussed to this point employ the same basic strategy:

128


1. Choose a direction S(x) that is designed to produce descent. 2. Exploit this direction to the maximum extent; that is, perform a line search. When we can no longer find a descent direction (as discussed in Section 3.1 and shown in Figure 3.2), we have located a local minimum (at least). Even direct-search methods like Hooke–Jeeves and simplex search use this basic strategy, even though it may not be explicit or obvious in the declaration of the algorithm. Trust region methods [21, 66, 71] use a slightly different strategy: 1. Form a ‘‘trustworthy’’ approximation of ƒ(x). 2. Find the minimum of the approximation. The ‘‘trust region’’ is defined as a neighborhood around the current iterate (x (t)) in which the current approximation of ƒ(x) produces descent. The emphasis now shifts from finding descent directions to forming ‘‘safe’’ approximations to ƒ(x). Marquardt’s method (see Section 3.3.4) is a trust region method. Notice that Marquardt’s method does not require a line search in order to converge. The choice of the parameter ␭ determines the approximation to be employed at each iteration, and the approximation becomes quadratic in the limit. Marquardt’s method also demonstrates the major problem with trust region methods, that is, the computational expense of forming the approximating function. 3.3.8

Gradient-Based Algorithm

The similarities of the conjugate gradient and quasi-Newton methods suggest that we might devise a general algorithm including all of the methods previously discussed. Such an algorithm is the gradient method: Gradient Method

Step 1. Define M ⫽ maximum number of allowable iterations N ⫽ number of variables x(0) ⫽ initial estimate of x* ␧1 ⫽ overall convergence criteria ␧2 ⫽ line search convergence criteria Step 2. Set k ⫽ 0. Step 3. Calculate ⵜƒ(x (k)). Step 4. Is 㛳ⵜƒ(x (k)㛳 ⭐ ␧1? Yes: Print ‘‘convergence: gradient’’; go to 13. No: Continue.

3.3


129

Step 5. Is k ⬎ M? Yes: Print ‘‘termination: k ⫽ M’’; go to 13. No: Continue. Step 6. Calculate s(x (k)). Step 7. Is ⵜƒ(x (k))s(x (k)) ⬍ 0? Yes: Go to 9. No: Set s(x (k)) ⫽ ⫺ⵜƒ(x (k)). Print ‘‘restart: bad direction’’; Go to 9. Step 8. Find ␣(k) such that ƒ(x (k) ⫹ ␣ (k)s(x (k))) → minimum using ␧2. Step 9. Set x (k⫹1) ⫽ x (k) ⫹ ␣ (k)s(x (k)). Step 10. Is ƒ(x (k⫹1)) ⬍ ƒ(x (k))? Yes: Go to 11. No: Print ‘‘termination: no descent’’; go to 13. Step 11. Is 㛳⌬ x㛳 / 㛳 x (k)㛳 ⭐ ␧1? Yes: Print ‘‘termination: no progress’’; go to 13. No: Go to 12. Step 12. Set k ⫽ k ⫹ 1. Go to 3. Step 13. Stop. We produce the various methods by defining the appropriate search direction in step 6 of the algorithm. The required calculation for each method has been given previously. Admittedly, some of the calculations (e.g., line 7) would not be appropriate for the Cauchy method, but we have discussed an algorithm for this method before. A purist might insist upon a restart feature for the FR method, but the checks included in the algorithm would catch any difficulties associated with the lack of restart for a conjugate gradient method. We might insert additional tests in line 7 similar to those suggested by Beale and Powell for restart. An efficient implementation would certainly contain additional tests in line 8 along the lines suggested by Davidon, Powell, and Shanno. Certainly, exact line searches should be avoided whenever possible. Our tests indicate that line searching is typically the most time consuming phase of these methods. 3.3.9

Numerical Gradient Approximations

To this point in our discussion of gradient methods we have assumed that the gradient ⵜƒ(x) and the Hessian matrix ⵜ2ƒ(x) are explicitly available. Modern methods seek to mimic the positive characteristics of Newton’s method, such as rate of convergence near the minimum, without use of second-order information [ⵜ2ƒ(x)]. In many practical problems the gradient is not explicitly available or is difficult and inconvenient to formulate. An example is when ƒ(x) is the result of a simulation. In addition, even when the analytical expressions can be obtained, they are prone to error. Therefore, if for no other

130


reason than to check the analytical expressions, it seems prudent to have a procedure to approximate the gradient numerically. We have used with considerable success the simple forward-difference approximation:

冏

⭸ƒ(x) ⭸xi

⫽

x⫽x

ƒ(x ⫹ ␧e(i)) ⫺ ƒ(x) ␧

(3.90)

This approximation follows directly from the definition of a partial derivative and is a good estimate for appropriately small values of ␧. The proper choice for ␧ depends on ƒ(x) and x and the accuracy or word length of the computing machine. We recall that theoretically the approximation is exact in the limit as ␧ goes to zero. This is obviously an unacceptable choice for ␧, and in fact ␧ must be large enough so that the difference in the number is finite. When ␧ is less than the precision of the machine, the numerator will be zero. In addition, we must not choose ␧ too large or lose contact with any semblance of a limiting process. In other words, for ␧ too large we get good numbers with regard to machine precision but poor estimates of the derivative. We can, at the expense of an additional function evaluation, increase the inherent accuracy of the approximation by using the central-difference formula:

冏

⭸ƒ(x) ⭸xi

x⫽x

⫽

ƒ(x ⫹ ␧e(i)) ⫺ ƒ(x ⫺ ␧e(i)) 2␧

(3.91)

For a given machine, ƒ(x), x, and ␧, the central-difference approximation will be more accurate but requires an additional function value. In many cases the increased accuracy does not warrant the additional expense. Stewart [72] proposed a procedure for calculating ␧ at each iteration of the DFP method. His method is based upon estimates of the inherent error of the forward-difference approximation (3.90) and the error introduced by subtracting nearly equal numbers in the numerator when ␧ becomes very small. The method is attractive because it requires no additional function values. Gill et al. [73] have shown that the method may produce poor values for ␧ and suggest a different approach. Their method requires 2N function values to calculate the required N␧ values but needs to be performed only one at x (0). Their technique is based upon better estimates of the errors involved, and numerical results suggest superiority of the Gill approach.

3.4

COMPARISON OF METHODS AND NUMERICAL RESULTS

Only recently has it been possible to say much from a theoretical viewpoint concerning the convergence or rate of convergence of the methods discussed in the previous sections. Accordingly, much of what is known about the ef-

3.4


131

ficiency of these methods when applied to general functions has come from numerical experiments. Very often the publication of a new method is associated with very sparse comparative numerical results. Often the test problems are chosen to highlight the outstanding characteristics of the new method. This manner of presentation limits to a large degree the validity of the results. Himmelblau (see Chap. 5 of [74]) gives numerical results of quite a different nature. His results have increased significance since he is not a personal advocate of any single method. He characterizes each method (direct and gradient) according to robustness, number of function evaluations, and computer time to termination when applied to a family of test problems. Robustness is a measure of success in obtaining an optimal solution of given precision for a wide range of problems. Himmelblau concludes that the BFGS, DFP, and Powell’s direct-search methods are the superior methods. The reader is referred directly to Himmelblau for complete conclusions. A similar but less complete study is given by Sargent and Sebastian [75]. They give results for gradient-based methods, including the BFGS, DFP, and Fletcher-Reeves algorithms. They test the effect of line search termination criteria, restart frequency, positive definiteness for quasi-Newton methods, and gradient accuracy. Their results once again indicate the superiority of the quasi-Newton methods for general functions (in particular BFGS). They also note that computer word length affects the quasi-Newton methods more than the conjugate gradient methods. This leads to the possibility that the Fletcher– Reeves method might be superior on a small computer such as is sometimes used in a process control environment. Carpenter and Smith [76] have reported on the relative computational efficiency of Powell’s direct-search method, DFP, and Newton’s method for structural applications. They conclude that for the test problems considered the DFP and Newton’s methods are superior to Powell’s direct-search method, and DFP was superior to Newton’s method for large problems. More recently, rather extensive tests have been conducted by Shanno and Phua [43, 59, 60]. They report on numerical experiments with a variety of conjugate gradient and variable metric methods. Their results are difficult to summarize briefly (the reader is referred directly to the cited papers), but it is fair to say that the tests suggest a superiority of the BFGS method over other available methods. It is probably not appropriate at this point in our discussion to devote much attention to the theory of numerical experiments. On the other hand, we

Table 3.2

Termination Criteria

1. k ⬎ M 㛳⌬ x㛳 2. (k) ⭐ ␧1 㛳x 㛳

3. 㛳ⵜƒ(x (k))㛳 ⭐ ␧1 4. ⵜƒ(x(k)) s(x (k)) ⬎ 0 5. ƒ(x(k⫹1)) ⬎ ƒ(x (k))

132

a

1.25 ⫻ 10 / 1.382 / 4,066 5.91 ⫻ 10⫺6 / 0.57 / 805 2.39 ⫻ 10⫺8 / 0.133 / 656 3.6 ⫻ 10⫺8 / 0.161 / 740

1.10 ⫻ 10 / 7.47 / 38,424 3.24 ⫻ 10⫺6 / 0.18 / 988 2.45 ⫻ 10⫺8 / 0.173 / 977 5.6 ⫻ 10⫺8 / 0.169 / 932

8.30 ⫻ 10 / 1.074 / 3,570 5.96 ⫻ 10⫺5 / 0.109 / 370 6.6 ⫻ 10⫺8 / 0.115 / 331 2.1 ⫻ 10⫺8 / 0.115 / 315 ⫺8

Coggin

6.19 ⫻ 10 / 2.631 / 10,685 2.77 ⫻ 10⫺7 / 0.059 / 273 4.3 ⫻ 10⫺8 / 0.063 / 239 3.9 ⫻ 10⫺9 / 0.065 / 204

Cubic ⫺9

ƒ(x*) / time / NF, where NF ⫽ number of function evaluations required to terminate, time ⫽ seconds of CPU time to termination, and ƒ(x*) ⫽ ƒ(x) at termination.

⫺10

Cauchy FR DFP BFGS

a

Gold

⫺10

Half

Rosenbrock’s Function (x (0) ⴝ [ⴚ1.2, 1.0]T)

Method

Table 3.3

3.4


133

should mention that testing guidelines are now available [77] and that most of the results given in the literature prior to approximately 1977 should be used cautiously. Certainly ratings on the basis of number of function evaluations only can be misleading, as shown recently by Miele and Gonzalez [78]. We have performed some numerical experiments designed to demonstrate the relatively utility of the Cauchy, FR, DFP, and BFGS methods on a CDC6500 using single-precision arithmetic. The FR model is restarted every N ⫹ 1 steps, and the algorithms terminate when a condition in Table 3.2 is met. Table 3.3 gives the results for Rosenbrock’s function (Fig. 3.16): ƒ(x) ⫽ 100(x2 ⫺ x 12)2 ⫹ (1 ⫺ x1)2

(3.92)

which has been devised to be a specific challenge to gradient-based methods. The results with various line-searching strategies are given, and numerical approximations to the gradient were used throughout. Notice that the Cauchy/ half combination supplied the best value of ƒ at termination but at the most expense. The most efficient with regard to number of function evaluations is seen to be the BFGS/cubic combination. ‘‘Half’’ is the interval-halving

Figure 3.16. Rosenbrock’s function.

134


Table 3.4

Eason and Fenton Problem No. 10 (x(0) ⴝ [0.5, 0.5]T)

Method

Half

Gold

Coggin

Cubic

Cauchy FR DFP BFGS

1.744 / 0.299 / 1026a 1.744 / 0.066 / 249 1.744 / 0.056 / 232 1.744 / 0.055 / 226

1.744 / 0.219 / 688 1.744 / 0.053 / 199 1.744 / 0.051 / 184 1.744 / 0.051 / 176

1.744 / 181 / 312 1.744 / 12.21 / 28141 1.744 / 0.09 / 208 1.744 / 0.087 / 195

1.744 / 0.091 / 171 1.744 / 0.025 / 92 1.744 / 0.079 / 262 1.744 / 0.043 / 133

a

ƒ(x*) / time / N.F.

method, ‘‘gold’’ is golden section, ‘‘Coggin’’ is a particular version of quadratic interpolation [24], and the ‘‘cubic’’ method involves the fitting of a single cubic at each state. Table 3.4 contains the results for a more practical problem, the minimization of a gear train inertia [79]. The problem (see Figure 3.17) is one of a set used by Eason and Fenton to compare various algorithms. The objective function is

Figure 3.17. Eason and Fenton’s gear inertia problem.

135

a

ƒ(x*) / time / NF.

1.77 ⫻ 10 / 10.70 / 29,692 1.2 ⫻ 10⫺8 / 0.311 / 942 3.7 ⫻ 10⫺9 / 0.684 / 1895 2.0 ⫻ 10⫺9 / 0.652 / 1650

3.74 ⫻ 10 / 13.466 / 39,503 1.3 ⫻ 10⫺10 / 0.401 / 1257 3.9 ⫻ 10⫺9 / 0.810 / 2404 2.2 ⫻ 10⫺9 / 0.759 / 2054 ⫺8

Cauchy FR DFP BFGS

a

Gold

⫺9

Half

Wood’s Function (x(0) ⴝ [ⴚ3, ⴚ1, ⴚ3, ⴚ1]T)

Method

Table 3.5

2.95 ⫻ 10 / 5.844 / 12,392 2.9 ⫻ 10⫺8 / 0.571 / 1468 2.9 ⫻ 10⫺9 / 0.403 / 747 2.0 ⫻ 10⫺9 / 0.444 / 723 ⫺9

Coggin

8 / 36 ⫻ 10⫺6 / 7.89 / 20,007 2.0 ⫻ 10⫺7 / 0.083 / 244 9.5 ⫻ 10⫺10 / 0.298 / 727 2.3 ⫻ 10⫺9 / 0.240 / 410

Cubic

136


冉

ƒ(x) ⫽ 12 ⫹ x 21 ⫹

冊冉冊

1 ⫹ x 22 x 21x 22 ⫹ 100 ⫹ x 12 (x1x2)4

1 10

(3.93)

On this problem we note that the FR/cubic algorithm is most efficient in both time and number of function evaluations. Table 3.5 gives the final group of results for Wood’s function [80] (see Figure 3.18): ƒ(x) ⫽ 100(x2 ⫺ x 21)2 ⫹ (1 ⫺ x1)2 ⫹ 90(x4 ⫺ x 23)2 ⫹ (1 ⫺ x3)2 ⫹ 10.1[(x2 ⫺ 1)2 ⫹ (x4 ⫺ 1)2] ⫹ 19.8(x2 ⫺ 1)(x4 ⫺ 1)

(3.94)

and once again the FR/cubic seems to be most efficient. These tests have been carefully executed, but we see that the results do not agree with other published results. We have found the FR method to be very efficient. Additional numerical testing by the reader is suggested.

Figure 3.18. Wood’s function.

REFERENCES

3.5

137

SUMMARY

In this chapter we have covered pertinent issues related to functions of several variables as they appear in an optimization setting. We have given the necessary and sufficient conditions for existence of a minimum of a function of several variables. On the other hand, most of the chapter is devoted to an orderly survey of methods. Some methods have been included because of their historical importance, while others represent the best that is currently available. Methods that require values of ƒ(x) only; ƒ(x) and ⵜƒ(x); and ƒ(x), ⵜƒ(x), and ⵜ2ƒ(x) are included in the discussions. The important concept of conjugacy is considered in some depth, and the latest conjugate gradient and quasi-Newton methods are cited. The discussion is not exhaustive, and many methods are omitted due to lack of space. The chapter closes with algorithmic considerations and brief numerical results.

REFERENCES 1. Murray, W. (Ed.), Numerical Methods for Unconstrained Optimization, Academic, London, 1972. 2. Nocedal, J., ‘‘Theory of Algorithms for Unconstrained Optimization,’’ Acta Numerica, 199–242 (1992). 3. Fletcher, R., Practical Methods of Optimization, 2nd ed., Wiley, New York, 2000. 4. Box, M. J., D. Davies, and W. H. Swann, Non-Linear Optimization Techniques, ICI Monograph 5, Oliver and Boyd, Edinburgh, 1969. 5. Krolak, P., and L. Cooper, ‘‘An Extension of Fibonaccian Search to Several Variables,’’ J. ACM, 6, 639–641 (1963). 6. Box, G. E. P., ‘‘Evolutionary Operation: A Method for Increasing Industrial Productivity,’’ Appl. Stat., 6, 81–101 (1957). 7. Spendley, W., G. R. Hext, and F. R. Himsworth, ‘‘Sequential Application of Simplex Designs in Optimization and Evolutionary Operation,’’ Technometrics, 4, 441–461 (1962). 8. Nelder, J. A., and R. Mean, ‘‘A Simplex Method for Function Minimization,’’ Computer J., 7, 308–313 (1965). 9. Box, G. E. P., and N. R. Draper, Evolutionary Operation, Wiley, New York, 1969. 10. Parkinson, J. M., and D. Hutchinson, ‘‘An Investigation into the Efficiency of Variants on the Simplex Method,’’ in Numerical Methods for Non-Linear Optimization (F. A. Lootsma, Ed.), Academic, London, 1972, pp. 115–135. 11. Powell, M. J. D., ‘‘On Search Directions for Minimization Algorithms,’’ Math. Prog., 4(2), 193–201 (1973). 12. Hooke, R., and T. A. Jeeves, ‘‘Direct Search of Numerical and Statistical Problems.’’ J. ACM, 8, 212–229 (1966). 13. Bandler, J. W., and P.A. McDonald, ‘‘Optimization of Microwave Networks by Razor Search,’’ IEEE Trans. Microwave Theory Tech., 17, 552–562 (1969).

138


14. Emery, F. E., and M. O. O’Hagan, ‘‘Optimal Design of Matching Networks for Microwave Transistor Amplifiers,’’ IEEE Trans. Microwave Theory Tech., 14, 696– 698 (1966). 15. Rosenbrock, H. H., ‘‘An Automated Method for Finding the Greatest or Least Value of a Function,’’ Computer J., 3(3), 175–184 (1960). 16. Swann, W. H. ‘‘Report on the Development of a New Direct Search Method of Optimization,’’ Central Instr. Res. Lab., Res. Note 64 / 3, Imperial Chemical Industries, Widnes, England, 1964. 17. Powell, M. J. D., ‘‘An Efficient Method for Finding the Minimum of a Function of Several Variables without Calculating Derivatives,’’ Computer J., 7, 155–162 (1964). 18. Zangwill, W. I., ‘‘Minimizing a Function without Calculating Derivatives,’’ Computer J., 10, 293–296 (1967). 19. Brent, R. P., Algorithms for Minimization without Derivatives, Prentice-Hall, Englewood Cliffs, NJ, 1973. 20. Gill, P. E., W. Murray, and M. H. Wright, Practical Optimization, Academic, New York, 1981, p. 57. 21. Fletcher, R., Practical Methods of Optimization, Vol. 1: Unconstrained Optimization, Wiley, New York, 1980, p. 17. 22. Fletcher, R., Practical Methods of Optimization, 2nd ed., Wiley, New York, 2000. 23. Cauchy. A., ‘‘Method Generale pour la Resolution des Systemes d’Equations Simultanees,’’ Compt. Rend. Acad. Sci., 25, 536–538 (1847). 24. Coggins, G. F., ‘‘Univariate Search Methods,’’ Central Instr. Res. Lab., Res. Note 64 / 11, Imperial Chemical Industries, Widnes, England, 1964. 25. Gabriele, G. A., and K. M. Ragsdell, OPTLIB: An Optimization Program Library, Purdue Research Foundation, West Lafayette, IN, Jan. 1977. 26. Ragsdell, K. M., R. R. Root, and G. A. Gabriele, ‘‘Newton’s Method: A Classical Technique of Contemporary Value,’’ ASME Design Technol. Transfer Conf. Proc., New York, Oct. 1974, p. 137. 27. Mangasarian, O. L., ‘‘Techniques of Optimization,’’ ASME paper 71-vibr.-118 presented at the ASME Design Automat. Conference, Sept. 8–10, 1971. 28. Marquardt, D. W., ‘‘An Algorithm for Least Squares Estimation of Non-Linear Parameters,’’ SIAM J., 11, 431–441 (1963). 29. Powell, M. J. D., ‘‘Problems Related to Unconstrained Optimization,’’ in Numerical Methods for Unconstrained Optimization, Academic, New York, 1972, Chap. 3, p. 29. 30. Bard, Y., Nonlinear Parameter Estimation, Academic, New York, 1974. 31. Hestenes, M. R., and E. Stiefel, ‘‘Methods of Conjugate Gradients for Solving Linear Systems,’’ NBS Res. J., 49, 409–436 (1952). 32. Fletcher, R., and C. M. Reeves, ‘‘Function Minimization by Conjugate Gradients,’’ Computer J., 7, 149–154 (1964). 33. Fried, I., and J. Metzler, ‘‘SOR vs Conjugate Gradients in a Finite Element Discretization,’’ Int. J. Numer. Methods Eng., 12, 1329–1342, (1978). 34. Miele, A., and J. W. Cantrell, ‘‘Study on a Memory Gradient Method for Minimization of Functions,’’ JOTA, 3(6), 459 (1969). 35. Polak, E., and G. Ribiere, ‘‘Note sur la Convergence de Methods de Directions Conjugees,’’ Rev. Fr. Inform. Rech. Operat., 16, 35–43 (1969).

REFERENCES

139

36. Davison, E. J., and P. Wong, ‘‘A Robust Conjugate-Gradient Algorithm Which Minimizes L-Functions,’’ Automatica, 11, 297–308 (1975). 37. Boland, W. R., E. R. Kamgnia, and J. S. Kowalik, ‘‘A Conjugate Gradient Optimization Method Invariant to Nonlinear Scaling,’’ JOTA, 27(2), 221–230 (1979). 38. Boland, W. R., and J. S. Kowalik, ‘‘Extended Conjugate-Gradient Methods with Restarts,’’ JOTA, 28(1), 1–9 (1979). 39. Crowder, H. P., and P. Wolfe, ‘‘Linear Convergence of the Conjugate Gradient Method,’’ IBM J. Res. Develop., 16, 431–433 (1972). 40. Powell, M. J. D., ‘‘Some Convergence Properties of the Conjugate Gradient Method,’’ Math. Prog., 11, 42–49 (1976). 41. Powell, M. J. D., ‘‘Restart Procedures for the Conjugate Gradient Method,’’ Math. Prog., 12, 241–254 (1977). 42. Beale, E. M. L., ‘‘A Derivation of Conjugate Gradients,’’ in Numerical Methods for Non-Linear Optimization (F. A. Lootsma, Ed.), Academic, New York, 1972, pp. 39–43. 43. Shanno, D. F., ‘‘Conjugate Gradient Methods with Inexact Searches,’’ Math. Oper. Res., 3(3), 244–256 (1978). 44. Gabriele, G. A., and K. M. Ragsdell, ‘‘Large Scale Nonlinear Programming Using the Generalized Reduced Gradient Method,’’ ASME J. Mech. Des., 102(3), 566– 573 (1980). 45. Avriel, M., Nonlinear Programming: Analysis and Methods, Prentice-Hall, Englewood Cliffs, NJ, 1976, p. 322. 46. Davidon, W. C., ‘‘Variable Metric Method for Minimization,’’ AEC Res. Develop. Rep., ANL-599, 1959. 47. Fletcher, R., and M. J. D. Powell, ‘‘A Rapidly Convergent Descent Method for Minimization,’’ Computer J., 6, 163 (1963). 48. Broyden, G. G., ‘‘The Convergence of a Class of Double-Rank Minimization Algorithms,’’ J. Inst. Math. Appl., 6, 76–90, 222–231 (1970). 49. Fletcher, R., ‘‘A New Approach to Variable Metric Algorithms,’’ Computer J., 13, 317–322 (1970). 50. Goldfarb, D., ‘‘Generating Conjugate Directions without Line Searches Using Factorized Variable Metric Updating Formulas,’’ Math Prog., 13, 94–110 (1977). 51. Shanno, D. F., ‘‘Conditioning of Quasi-Newton Methods for Function Minimization,’’ Math. Comp., 24, 647–657 (1970). 52. Powell, M. J. D., ‘‘A View of Unconstrained Optimization,’’ paper presented at IMA Conference, ‘‘Optimization in Action,’’ Jan. 1975. 53. Davidon, W. C., ‘‘Optimally Conditioned Optimization Algorithms without Line Searches,’’ Math. Prog., 9, 1–30 (1975). 54. Powell, M. J. D., ‘‘Quadratic Termination Properties of Davidon’s New Variable Metric Algorithm,’’ Math. Prog., 12, 141–147 (1977). 55. McCormick, G. P., ‘‘Methods of Conjugate Directions versus Quasi-Newton Methods,’’ Math. Prog., 3(1), 101–116 (1972). 56. Shanno, D. F., and K. H. Phua, ‘‘A Variable Method Subroutine for Unconstrained Nonlinear Minimization,’’ Manag. Inform. Syst. Rep. 28, University of Arizona, Tucson, AZ, 1979. 57. Dixon, L. C. W., ‘‘Quasi-Newton Algorithms Generate Identical Points,’’ Math. Prog., 2(3), 383–387 (1972).

140


58. Dixon, L. C. W., ‘‘Quasi-Newton Algorithms Generate Identical Points II: The Proofs of Four New Theorems,’’ Math. Prog., 3(3), 345–358 (1972). 59. Shanno, D. F., and K. H. Phua, ‘‘Matrix Conditioning and Nonlinear Optimization,’’ Math. Prog., 14, 149–160 (1978). 60. Shanno, D. F., and K. H. Phua, ‘‘Numerical Comparison of Several VariableMetric Algorithms,’’ JOTA, 25(4), 507–518 (1978). 61. Shanno, D. F., and K. H. Phua, ‘‘Algorithms 500: Minimization of Unconstrained Multivariate Functions [E4], ‘‘ACM Trans. Math. Software, 2(1), 87–94 (1976). 62. LaFrance, L. J., J. F. Hamilton, and K. M. Ragsdell, ‘‘On the Performance of a Combined Generalized Reduced Gradient and Davidon-Fletcher-Powell Algorithm: GRGDFP,’’ Eng. Opt., 2, 269–278 (1977). 63. Root, R. R., and K. M. Ragsdell, ‘‘Computational Enhancements to the Method of Multipliers,’’ ASME J. Mech. Des., 102(3), 517–523 (1980). 64. Han, S. P., ‘‘Superlinearly Convergent Variable Metric Algorithms for General Nonlinear Programming Problems,’’ Math. Prog., 11, 263–282 (1976). 65. Powell, M. J. D., ‘‘The Convergence of Variable Metric Method for Nonlinear Constrained Optimization Calculations,’’ in Nonlinear Programming, Vol. 3 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, Eds.), Academic, New York, 1978, pp. 27–64. 66. Nocedal, J., and S. J. Wright, Numerical Optimization, Springer Science, New York, 1999. 67. Toint, Ph. L., ‘‘On Sparse and Symmetric Matrix Updating Subject to a Linear Equation,’’ Math. Comp., 31(140), 954–961 (1977). 68. Powell, M. J. D., and Ph. L. Toint, ‘‘On the Estimation of Sparse Hessian Matrices,’’ SIAM J. Numer. Anal., 16, 1060–1074 (1979). 69. Shanno, D. F., ‘‘On Variable-Metric Methods for Sparse Hessians,’’ Math. Comp., 34(150), 499–514 (1980). 70. Shanno, D. F., ‘‘Computational Experience with Methods for Estimating Sparse Hessians for Nonlinear Optimization,’’ JOTA, 35(2), 183–193 (Oct. 1981). 71. Sorensen, D. C., ‘‘Trust Region Methods for Unconstrained Optimization,’’ in Nonlinear Optimization—1981 (M. J. D. Powell, Ed.), Academic Press, London, 1981, p. 29. 72. Stewart, G. W., ‘‘A Modification of Davidon’s Method to Accept Difference Approximations of Derivatives,’’ JACM, 14, 72–83 (1967). 73. Gill, P. E., W. Murray, M. A. Saunders, and M. H. Wright, ‘‘Computing FiniteDifference Approximations to Derivatives for Numerical Optimization,’’ Stanford Systems Optimization Lab. Tech. Rep. SOL80-6, Palo Alto, CA, 1980. 74. Himmelblau, D. M., Applied Nonlinear Programming, McGraw-Hill, New York, 1972, p. 190. 75. Sargent, R. W. H., and D. J. Sebastian, ‘‘Numerical Experience with Algorithms for Unconstrained Minimization,’’ in Numerical Methods for Non-Linear Optimization (F. A. Lootsma, Ed.), Academic, New York, 1971, Chap. 5, p. 45. 76. Carpenter, W. C., and E. A. Smith, ‘‘Computational Efficiency in Structural Optimization,’’ Eng. Optimization, 1(3), 169–188 (1975). 77. Crowder, H., R. S. Dembo, and J. M. Mulvey, ‘‘On Reporting Computational Experiments with Mathematical Software,’’ ACM Trans. Math. Software, 5(2), 198–203 (1979).

PROBLEMS

141

78. Miele, A., and S. Gonzalez, ‘‘On the Comparative Evaluation of Algorithms for Mathematical Programming Problems,’’ in Nonlinear Programming, Vol. 3 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, Eds.), Academic, New York, 1978, pp. 337–359. 79. Eason, E. D., and R. G. Fenton, ‘‘A Comparison of Numerical Optimization Methods for Engineering Design,’’ J. Eng. Ind. Trans. ASME, 96(1), 196–200 (1974). 80. Colville, A. R., ‘‘A Comparative Study of Nonlinear Programming Codes,’’ Tech. Rep. 320-2949, IBM New York Scientific Center, 1968. 81. Sherwood, T. K., A Course in Process Design, MIT Press, Cambridge, MA, 1963, p. 84.

PROBLEMS

3.1. Explain why the search directions used in a direct-search algorithm such as Hooke–Jeeves must be linearly independent. How many direction vectors must be used? 3.2. Cite two situations under which the simplex search method would be preferable to Powell’s conjugate direction method. 3.3. Why do we use quadratic functions as the basis for developing nonlinear optimization algorithms? 3.4. What is the utility of the parallel subspace property of quadratic functions? 3.5. What is the descent property? Why is it necessary? Cite one algorithm which does have and one which does not have this property. 3.6. Why is it necessary for the Ak matrix of quasi-Newton methods to be positive definite in a minimization problem? 3.7. Explain the relationship of Marquardt’s method to the Cauchy and Newton methods. Of the three, which is preferred? 3.8. Is it possible that DFP and Fletcher–Reeves methods generate identical points if both are started at the same point with a quadratic function? If so, under what conditions? If not, why not? 3.9. Explain quadratic convergence. Cite one algorithm which does have and one which does not have this property. 3.10. Show that the function ƒ(x) ⫽ 3x 21 ⫹ 2x 22 ⫹ x 23 ⫺ 2x1x2 ⫺ 2x1x3 ⫹ 2x2x3 ⫺ 6x1 ⫺ 4x2 ⫺ 2x3 is convex.

142


3.11. Find and classify the stationary points of ƒ(x) ⫽ 2x 31 ⫹ 4x1x 22 ⫺ 10x1x2 ⫹ x 22 as shown in Figure 3.19. 3.12. Investigate the definiteness of the following quadratic forms: Q1(x) ⫽ x 21 ⫹ 2x 22 ⫺ 3x 32 ⫺ 6x1x2 ⫹ 8x1x3 ⫺ 4x2 x3 Q2(x) ⫽ 2ax1x2 ⫹ 2bx2x3 ⫹ 2cx3x1 Q3(x) ⫽ x 21 ⫹ 5x 22 ⫹ 3x 32 ⫹ 4x1x2 ⫺ 2x2x3 ⫺ 2x1x3 3.13. Use Gaussian elimination to ‘‘complete the square’’ and show that Q(x) ⫽ x 21 ⫹ 2x1x2 ⫹ 4x1x3 ⫹ 3x 22 ⫹ 2x2x3 ⫹ 5x 23 is a positive-definite quadratic form.

Figure 3.19. Function of problem 3.11.

PROBLEMS

143

3.14. Suppose at x ⫽ x, ⵜƒ(x) ⫽ 0. What can you say about x if (a) ƒ(x) is convex? (b) ƒ(x) is concave? (c) ⵜ2ƒ(x) is indefinite? (d) ⵜ2ƒ(x) is positive semidefinite? (e) ⵜ2ƒ(x) is negative semidefinite? 3.15. Peano’s Counterexample. Consider the function ƒ(x) ⫽ (x1 ⫺ a 12x 22)(x1 ⫺ a 22x 22) with a1 and a2 constant. (a) Classify the point x ⫽ [0, 0]T. (b) Show that ƒ(x) attains a maximum at the origin for all points on the curve x1 ⫽ –12 (a 21 ⫹ a 22)x 22 (c) Draw several contours of the function in the neighborhood of the origin. 3.16. While searching for the minimum of ƒ(x) ⫽ [x 21 ⫹ (x2 ⫹ 1)2][x 12 ⫹ (x2 ⫺ 1)2] we terminate at the following points: (a) x (1) ⫽ [0, 0]T (b) x (2) ⫽ [0, 1]T (c) x (3) ⫽ [0, ⫺1]T (d) x (4) ⫽ [1, 1]T Classify each point. 3.17. Suppose you are a fill-dirt contractor and you have been hired to transport 400 yd3 of material across a large river. To transport the material, a container must be constructed. The following cost data are available: Each round trip costs $4.20 and material costs for the container are Bottom Sides Ends

20.00/yd2 5.00/yd2 20.00/yd2

Design a container to minimize cost that will perform the assigned task.

144


3.18. Consider Rosenbrock’s function: ƒ(x) ⫽ 100(x2 ⫺ x 12)2 ⫹ (1 ⫺ x1)2 and the starting point x (0) ⫽ [⫺1.2, 0]T Find x* such that ƒ(x*) is a minimum, (a) Using the Nelder and Mead simplex search method, perform four iterations by hand, and then using x (0) again use SPX in OPTLIB [25] or other appropriate software of your choice. (b) Using the Hooke–Jeeves method (PS in OPTLIB). (c) Using Powell’s conjugate directions method (PCD in OPTLIB). 3.19. Consider the grain storage bin given in Figure 3.20. Choose h, d, and ␾ such that the bin holds a prescribed volume (v* ⫽ 10 m3) and is of minimum cost. Assume the base is made of wood at C1 ⫽ $1/m2, and the remainder is constructed of sheet metal at C2 ⫽ $1.5/m2. Use the volume constraint to eliminate one variable, and solve the resulting twovariable problem using the pattern search method (␪ ⫽ 30⬚). Hint: You may find it useful to describe the bin with another (but equivalent) set of design variables. 3.20. Consider a gas pipeline transmission system as shown in Figure 3.21 (Sherwood [81]), where compressor stations are placed L miles apart. Assume that the total annual cost of the transmission system and its operation is

Figure 3.20. Grain storage bin.

PROBLEMS

145

Figure 3.21. Gas pipeline.

C(D, P1, L, r) ⫽ 7.84D2P1 ⫹ 450,000 ⫹ 36,900D ⫹ ⫹

where D P1 L r

⫽ ⫽ ⫽ ⫽

(772)(106) 0.219 (r ⫺ 1) L

(6.57)(106) L

($/yr)

(1)

pipe ID, in. compressor discharge pressure, psia length between compressor stations, miles compression ratio ⫽ P1 /P2

Furthermore, assume that the flow rate is

冋

Q ⫽ 3.39

(P 21 ⫺ P 22)D5 ƒL

册

1/2

(SCF/hr)

(2)

with ƒ ⫽ friction factor ⫽ 0.008D⫺1 / 3. Let the flow rate be 100 ⫻ 106 SCF/day, and use expression (2) to eliminate P1 from (1). Then find the minimum-cost design using the simplex search method of Nelder and Mead and Powell’s conjugate directions direct-search method. 3.21. Locate the minima of Himmelblau’s function: ƒ(x) ⫽ (x 21 ⫹ x2 ⫺ 11)2 ⫹ (x1 ⫹ x 22 ⫺ 7)2 (see Figure 3.1) from x (1) ⫽ [5, 5]T

x (2) ⫽ [5, ⫺5]T

x (4) ⫽ [⫺5, ⫺5]T

x (3) ⫽ [0, 0]T

x (5) ⫽ [5, 0]T

using the pattern search method of Hooke–Jeeves. Terminate when the elements of x are accurate to three decimal places. 3.22. Given the current iterate x (k), search direction S(x (k)), and an iteration of the form in (3.42), show that

146


␣* ⫽

ⵜq(x (k))TS(x (k))

s(x (k))TCS(x (k))

assuming the objective function is a quadratic: ƒ(x) ⫽ q(x) ⫽ a ⫹ bTx ⫹ –12 xTCx 3.23. Find the dimensions of the minimum-cost open-top rectangular container. (Let v* ⫽ 10 m3.) 3.24. Given q(x) ⫽ 8x 21 ⫹ 4x1x2 ⫹ 5x 22, x (0) ⫽ [10, 10]T, and two linearly independent directions d (0) ⫽ ⵜq(x (0)) ⫽ [200, 140]T

d (1) ⫽ [7, ⫺10]T

Form a new direction (S (1) ⫽ d (1) ⫹ ␤11 ⌬g (1)) conjugate to d (0). Search them both in turn to find x*. That is, search along S (0) ⫽ d (0) first; then, from this minimum point, search along S (1). How does S (1) compare to the Fletcher–Reeves direction? 3.25. Find a direction orthogonal to the vector s1 ⫽

冋

册

1 1 1 ,⫺ ,⫺ 兹3 兹3 兹3

T

x ⫽ [0, 0, 0]T

Also find a new direction s2 conjugate to s1 with respect to the objective function ƒ(x) ⫽ x1 ⫹ 2x 22 ⫺ x1x2 ⫹ 3x 21 ⫺ 2x1x3 ⫹ x 23 starting from the point x. 3.26. Locate and classify the stationary points of ƒ(x) ⫽ x 31 ⫺ x1x2 ⫹ x 22 ⫺ 2x1 ⫹ 3x2 ⫺ 4 3.27. Eason has shown that the allocation of gear ratios in a triple-reduction spur-gear train for minimum inertia with an overall reduction of 10 reduces to the minimization of the function in (3.93). Use the Hooke– Jeeves, Cauchy, and Fletcher–Reeves methods to approximate x* with x (0) ⫽ [0.5, 0.5]T. 3.28. Given the function ƒ(x) ⫽ 100(x2 ⫺ x 12)2 ⫹ (1 ⫺ x1)2 and x (0) ⫽ [⫺1.2, 1]T

x (1) ⫽ [⫺1.3, 1.07]T

PROBLEMS

147

the first two points in a search for x*, the minimum of ƒ. Calculate the search direction at x (1) using the following gradient-based methods: (a) Cauchy (b) Modified Newton (c) Fletcher–Reeves (d) Marquardt (␭(1) ⫽ 100) 3.29. We would like to study the effect of ␧2, the line search convergence criterion, on the performance of the Fletcher–Reeves, Davidon– Fletcher–Powell, and Broyden–Fletcher–Goldfarb–Shanno methods. Apply each method to Wood’s function (3.94) from x (0) ⫽ [⫺3, ⫺1, ⫺3, ⫺1]T. Let ␧2 ⫽ ␣␧1, ␣ ⫽ 0.01, 0.1, 1, 10, and ␧1, the overall convergence criterion, chosen appropriately for the computing environment. Note: On a CDC-6600, ␧1 ⫽ 10⫺4 is appropriate. 3.30. Consider once again the grain storage bin of problem 3.19. Investigate the effect of the relative cost c2 /c1 on the shape of the optimal bin. In particular, find the optimal design when c2 /c1 ⫽ 0.5, 1.5, 3. Draw the resulting designs. What general conclusions can be drawn? 3.31. Perform three iterations of the Cauchy, modified Newton, and DFP methods on Powell’s function: ƒ(x) ⫽ (x1 ⫹ 10x2)2 ⫹ 5(x3 ⫺ x4)2 ⫹ (x2 ⫺ 2x3)4 ⫹ 10(x1 ⫺ x4)4

from x (0) ⫽ [3, ⫺1, 0, 1]T 3.32. An engineer has performed some experiments and observed a certain quantity Q as a function of a variable t. He has good reasons to believe that there is a physical law relating t and Q that takes the form Q(t) ⫽ a sin t ⫹ b tan t ⫹ c He wants to have the ‘‘best’’ possible idea of the value of the coefficients a, b, and c using the results of his n experiments (t1, Q1; t2, Q2; t3, Q3; . . . ; tn, Qn). Taking into account that his experiments are not perfect and also perhaps that the relation is not rigorous, he does not expect to find a perfect fit. So he defines an error term ei, where ei ⫽ Qi ⫺ Q(ti) His first idea is to minimize the function

148


冘e n

R⫽

2 i

i⫽1

Show with detailed justifications that he will have to solve a system of linear equations in order to find the best values of a, b, and c. Hint: First show that the engineers’ problem reduces to minimizing an unconstrained convex function. 3.33. From theoretical considerations it is believed that the dependent variable y is related to variable x via a two-parameter function: y(x) ⫽

k1x 1 ⫹ k2 x

The parameters k1 and k2 are to be determined by a least-squares fit of the following experimental data: 1.0 2.0 3.0 4.0

1.05 1.25 1.55 1.59

Find k1 and k2 using the BFGS variable metric method. 3.34. Show that ␥ (i) ⫽ 0, i ⫽ 0, 1, 2, . . . , k ⫺ 2, when the conjugate gradient update form (3.62) is employed with S (0) ⫽ ⫺g (0), (3.42), and ƒ(x) is quadratic.

4 LINEAR PROGRAMMING The term linear programming defines a particular class of optimization problems in which the constraints of the system can be expressed as linear equations or inequalities and the objective function is a linear function of the design variables. Linear programming (LP) techniques are widely used to solve a number of military, economic, industrial, and societal problems. The primary reasons for its wide use are the availability of commercial software to solve very large problems and the ease with which data variation (sensitivity analysis) can be handled through LP models.

4.1

FORMULATION OF LINEAR PROGRAMMING MODELS

Formulation refers to the construction of LP models of real problems. Model building is not a science; it is primarily an art that is developed mainly by experience. The basic steps involved in formulating an LP model are to identify the design/decision variables, express the constraints of the problem as linear equations or inequalities, and write the objective function to be maximized or minimized as a linear function. We shall illustrate the basic steps in formulation with some examples. Example 4.1

An Inspection Problem

A company has two grades of inspectors, 1 and 2, who are to be assigned to a quality control inspection. It is required that at least 1800 pieces be inspected per 8-hr day. Grade 1 inspectors can check pieces at the rate of 25 per hour, with an accuracy of 98 percent. Grade 2 inspectors check at the rate of 15 pieces per hour, with an accuracy of 95 percent. Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

149

150

LINEAR PROGRAMMING

The wage rate of a grade 1 inspector is $4.00/hr, while that of a grade 2 inspector is $3.00/hr. Each time an error is made by an inspector, the cost to the company is $2.00. The company has available for the inspection job 8 grade 1 inspectors and 10 grade 2 inspectors. The company wants to determine the optimal assignment of inspectors that will minimize the total cost of the inspection. Formulation. Let x1 and x2 denote the number of grade 1 and grade 2 inspectors assigned to inspection. Since the number of available inspectors in each grade is limited, we have the following constraints: x1 ⭐ 8

(grade 1)

x2 ⭐ 10

(grade 2)

The company requires that at least 1800 pieces be inspected daily. Thus, we get 8(25)x1 ⫹ 8(15)x2 ⫽ 200x1 ⫹ 120x2 ⭓ 1800 or 5x1 ⫹ 3x2 ⭓ 45 To develop the objective function, we note that the company incurs two types of costs during inspection: (1) wages paid to the inspector and (2) cost of his inspection errors. The cost of each grade 1 inspector is $4 ⫹ $2(25)(0.02) ⫽ $5/hr Similarly, for each grade 2 inspector, $3 ⫹ $2(15)(0.05) ⫽ $4.50/hr Thus, the objective function is to minimize the daily cost of inspection given by Z ⫽ 8(5x1 ⫹ 4.50x2) ⫽ 40x1 ⫹ 36x2 The complete formulation of the LP problem thus becomes Minimize

Z ⫽ 40x1 ⫹ 36x2

Subject to x1 ⭐ 8

x2 ⭐ 10

5x1 ⫹ 3x2 ⭓ 45 x1 ⭓ 0

x2 ⭓ 0

4.1

Example 4.2


151

Hydroelectric Power Systems Planning

An agency controls the operation of a system consisting of two water reservoirs with one hydroelectric power generation plant attached to each, as shown in Figure 4.1. The planning horizon for the system is broken into two periods. When the reservoir is at full capacity, additional inflowing water is spilled over a spillway. In addition, water can also be released through a spillway as desired for flood protection purposes. Spilled water does not produce any electricity. Assume that on an average 1 kilo-acre-foot (KAF) of water is converted to 400 megawatt hours (MWh) of electricity by power plant A and 200 MWh by power plant B. The capacities of power plants A and B are 60,000 and 35,000 MWh per period. During each period, up to 50,000 MWh of electricity can be sold at $20.00/MWh, and excess power above 50,000 MWh can only be sold for $14.00/MWh. The following table gives additional data on the reservoir operation and inflow in kilo-acre-feet:

Capacity Predicted flow Period 1 Period 2 Minimum allowable level Level at the beginning of period 1

Reservoir A

Reservoir B

2000

1500

200 130 1200 1900

40 15 800 850

Develop a linear programming model for determining the optimal operating policy that will maximize the total revenue from electricity sales. Formulation. This example illustrates how certain nonlinear objective functions can be handled by LP methods. The nonlinearity in the objective function is due to the differences in the unit megawatt-hour revenue from electricity sales depending on the amount sold. The nonlinearity is apparent if a graph is plotted between total revenue and electricity sales. This is illustrated by the graph of Figure 4.2, which is a piecewise linear function, since it is linear in the regions (0, 50000) and (50000, ⬁). Hence, by partitioning

Figure 4.1. Hydroelectric system.

152

LINEAR PROGRAMMING

Figure 4.2. Electricity sales, MWh.

the quantity of electricity sold into two parts, the part that is sold for $20.00/ MWh and the part that is sold for $14.00/MWh, the objective function could be represented as a linear function. Let us first formulate the problem of reservoir operation for period 1 and then extend it to period 2. PH1 PL1 XA1 XB1 SA1 SB1 EA1 EB1

Power sold at $20/MWh, MWh Power sold at $14/MWh, MWh Water supplied to power plant A, KAF Water supplied to power plant B, KAF Spill water drained from reservoir A, KAF Spill water drained from reservoir B, KAF Reservoir A level at the end of period 1, KAF Reservoir B level at the end of period 1, KAF

Total power produced in period 1 ⫽ 400XA1 ⫹ 200XB1, while total power sold ⫽ PH1 ⫹ PL1. Thus, the power generation constraint becomes 400XA1 ⫹ 200XB1 ⫽ PH1 ⫹ PL1 Since the maximum power that can be sold at the higher rate is 50,000 MWh, we get PH1 ⭐ 50,000. The capacity of power plant A, expressed in water units (KAF), is 60,000 ⫽ 150 400 Hence, XA1 ⭐ 150 Similarly,

4.1


153

XB1 ⭐ 175 The conservation of water flow for reservoir A is given by Water supplied to power plant A ⫹ spill water ⫹ ending reservoir level ⫽ beginning reservoir level ⫹ predicted inflow

XA1 ⫹ SA1 ⫹ EA1 ⫽ 1900 ⫹ 200 ⫽ 2100 Since the capacity of reservoir A is 2000 KAF and the minimum allowable level is 1200, we get the following additional constraints: EA1 ⭐ 2000

EA1 ⭓ 1200

Similarly, for reservoir B, we get XB1 ⫹ SB1 ⫹ EB1 ⫽ 850 ⫹ 40 ⫹ XA1 ⫹ SA1 800 ⭐ EB1 ⭐ 1500 We can now develop the model for period 2, by defining the appropriate decision variables PH2, PL2, XA2, XB2, SA2, SB2, EA2, and EB2. The power generation constraints for period 2 are given by 400XA2 ⫹ 200XB2 ⫽ PH2 ⫹ PL2 PH2 ⭐ 50,000

XA2 ⭐ 150

XB2 ⭐ 175

The water flow constraints are given by XA2 ⫹ SA2 ⫹ EA2 ⫽ EA1 ⫹ 130 1200 ⭐ EA2 ⭐ 2000 XB2 ⫹ SB2 ⫹ EB2 ⫽ EB1 ⫹ 15 ⫹ XA2 ⫹ SA2 800 ⭐ EB2 ⭐ 1500 In addition to the above constraints, we have the nonnegativity restrictions on all the decision variables. The objective function is to maximize the total revenue from sales, given by Maximize

Z ⫽ 20(PH1 ⫹ PH2) ⫹ 14(PL1 ⫹ PL2)

Thus, the final model has 16 decision variables and 20 constraints, excluding the nonnegativity constraints.

154

LINEAR PROGRAMMING

4.2 GRAPHICAL SOLUTION OF LINEAR PROGRAMS IN TWO VARIABLES

In Section 4.1, some examples were presented to illustrate how practical problems can be formulated mathematically as LP problems. The next step after formulation is to solve the problem mathematically to obtain the best possible solution. In this section, a graphical procedure to solve LP problems involving only two variables is discussed. Although in practice such small problems are not usually encountered, the graphical procedure is presented to illustrate some of the basic concepts used in solving large LP problems. Example 4.3

Recall the inspection problem given by Example 4.1: Minimize

Z ⫽ 40x1 ⫹ 36x2

Subject to x1 ⭐ 8

x2 ⭐ 10

5x1 ⫹ 3x2 ⭓ 45 x1 ⭓ 0

x2 ⭓ 0

In this problem, we are interested in determining the values of the variables x1 and x2 that will satisfy all the restrictions and give the least value for the objective function. As a first step in solving this problem, we want to identify all possible values of x1 and x2 that are nonnegative and satisfy the constraints. For example, the point x1 ⫽ 8, x2 ⫽ 10 is positive and satisfies all the constraints. Such a point is called a feasible solution. The set of all feasible solutions is called the feasible region. Solution of a linear program is nothing but finding the best feasible solution in the feasible region. The best feasible solution is called an optimal solution to the LP problem. In our example, an optimal solution is a feasible solution that minimizes the objective function 40x1 ⫹ 36x2. The value of the objective function corresponding to an optimal solution is called the optimal value of the linear program. To represent the feasible region in a graph, every constraint is plotted, and all values of x1, x2 that will satisfy these constraints are identified. The nonnegativity constraints imply that all feasible values of the two variables will lie in the first quadrant. The constraint 5x1 ⫹ 3x2 ⭓ 45 requires that any feasible solution (x1, x2) to the problem should be on one side of the straight line 5x1 ⫹ 3x2 ⫽ 45. The proper side is found by testing whether the origin satisfies the constraints. The line 5x1 ⫹ 3x2 ⫽ 45 is first plotted by taking two convenient points (e.g., x1 ⫽ 0, x2 ⫽ 15 and x1 ⫽ 9, x2 ⫽ 0).

4.2

GRAPHICAL SOLUTION OF LINEAR PROGRAMS IN TWO VARIABLES

155

The proper side is indicated by an arrow directed above the line, since the origin does not satisfy the constraint. Similarly, the constraints x1 ⭐ 8 and x2 ⭐ 10 are plotted. The feasible region is given by the shaded region ABC shown in Figure 4.3. Obviously, there are an infinite number of feasible points in this region. Our objective is to identify the feasible point with the lowest value of Z. Observe that the objective function, given by Z ⫽ 40x1 ⫹ 36x2, represents a straight line if the value of Z is fixed a priori. Changing the value of Z essentially translates the entire line to another straight line parallel to itself. To determine an optimal solution, the objective function line is drawn for a convenient value of Z such that it passes though one or more points in the feasible region. Initially Z is chosen as 600. When this line is moved closer to the origin, the value of Z is further decreased (Figure 4.3). The only limitation on this decrease is that the straight line 40x1 ⫹ 36x2 ⫽ Z must contain at least one point in the feasible region ABC. This clearly occurs at the corner

Figure 4.3. Graphical solution of Example 4.3.

156

LINEAR PROGRAMMING

point A given by x1 ⫽ 8, x2 ⫽ –53 . This is the best feasible point, giving the lowest value of Z as 380. Hence, x1 ⫽ 8

x2 ⫽ –35

is an optimal solution, and Z ⫽ 380 is the optimal value for the linear program. Thus for the inspection problem the optimal utilization is achieved by using eight grade 1 inspectors and 1.6 grade 2 inspectors. The fractional value x2 ⫽ –53 suggests that one of the grade 2 inspectors is utilized for only 60 percent of the time. If this is not feasible, the normal practice is to round off the fractional values to get an approximate optimal integer solution as x1 ⫽ 8, x2 ⫽ 2. Unique Optimal Solution. In Example 4.3 the solution x1 ⫽ 8, x2 ⫽ –53 is the

only feasible point with the lowest value of Z. In other words, the values of Z corresponding to the other feasible solutions in Figure 4.3 exceed the optimal value of 380. Hence, for this problem, the solution x1 ⫽ 8, x2 ⫽ –53 is called a unique optimal solution. Alternative Optimal Solutions. In some LP problems there may exist more

than one feasible solution whose objective function values equal the optimal value of the linear program. In such cases, all these feasible solutions are optimal solutions, and the linear program is said to have alternative or multiple optimal solutions. Example 4.4

Maximize

Z ⫽ x1 ⫹ 2x2

Subject to x1 ⫹ 2x2 ⭐ 10 x1 ⫹ x2 ⭓ 1 x2 ⭐ 4 x1 ⭓ 0

x2 ⭓ 0

The feasible region is shown in Figure 4.4. The objective function lines are drawn for Z ⫽ 2, 6, 10. The optimal value for the linear program is 10, and the corresponding objective function line x1 ⫹ 2x2 ⫽ 10 coincides with side BC of the feasible region. Thus, the corner-point feasible solutions x1 ⫽ 10, x2 ⫽ 0 (B), x1 ⫽ 2, x2 ⫽ 4 (C), and all other feasible points on the line BC are optimal solutions.

4.2

GRAPHICAL SOLUTION OF LINEAR PROGRAMS IN TWO VARIABLES

157

Figure 4.4. Graphical solution of Example 4.4.

Unbounded Optimum. It is possible for some LP problems not to have an

optimal solution. In other words, it is possible to find better feasible solutions improving the objective function values continuously. This would have been the case if the constraint x1 ⫹ 2x2 ⭐ 10 were not given in Example 4.4. In this case, moving farther away from the origin increases the objective function x1 ⫹ 2x2, and maximum Z would be ⫹⬁. In cases where no finite maximum exists, the linear program is said to have an unbounded optimum. It is inconceivable for a practical problem to have an unbounded solution, since this implies that one can make infinite profit from a finite amount of resources! If such a solution is obtained in a practical problem, it usually means that one or more constraints have been omitted inadvertently during the initial formulation of the problem. These constraints would have prevented the objective function from assuming infinite values. Conclusion. In Example 4.3 the optimal solution was unique and occurred at the corner point A in the feasible region. In Example 4.4 we had multiple optimal solutions to the problem, which included two corner points B and C. In either case, one of the corner points of the feasible region was always an optimal solution. As a matter of fact, the following property is true for any LP problem.

158

LINEAR PROGRAMMING

Property 1 If there exists an optimal solution to a linear programming problem, then at least one of the corner points of the feasible region will always qualify to be an optimal solution. This is the fundamental property on which the simplex method for solving LP problems is based. Even though the feasible region of an LP problem contains an infinite number of points, an optimal solution can be determined by merely examining the finite number of corner points in the feasible region. For instance, in Example 4.4 the objective function was to maximize Z ⫽ x1 ⫹ 2x2 The corner points of the feasible region were A (1, 0), B (10, 0), C (2, 4), D (0, 4), and E (0, 1). Evaluating their Z values, we get Z(A) ⫽ 1, Z(B) ⫽ 10, Z(C) ⫽ 10, Z(D) ⫽ 8, and Z(E) ⫽ 2. Since the maximum value of Z occurs at corner points B and C, both are optimal solutions by Property 1. 4.3

LINEAR PROGRAM IN STANDARD FORM

The standard form of an LP problem with m constraints and n variables can be represented as follows: Maximize or minimize Z ⫽ c1x1 ⫹ c2x2 ⫹ ⫹ cn xn Subject to

a11x1 ⫹ a12x2 ⫹ ⫹ a1n xn ⫽ b1 a21x1 ⫹ a22x2 ⫹ ⫹ a2n xn ⫽ b2 ⯗

am1x1 ⫹ am2x2 ⫹ ⫹ amn xn ⫽ bm x1 ⭓ 0

x2 ⭓ 0

xn ⭓ 0

b1 ⭓ 0

b2 ⭓ 0 bm ⭓ 0

In matrix–vector notation, the standard LP problem can be expressed compactly as Maximize or minimize Z ⫽ cx Subject to

Ax ⫽ b x⭓0

b⭓0

where A is an m ⫻ n matrix, x is an n ⫻ 1 column vector, b is an m ⫻ 1 column vector, and c is a 1 ⫻ n row vector.

4.3

LINEAR PROGRAM IN STANDARD FORM

159

In practice, A is called the coefficient matrix, x is the decision vector, b is the requirement vector, and c is the profit (cost) vector of the linear program. The simplex method for solving LP problems requires that the problem be expressed in standard form. But not all linear programs come in standard form. Very often the constraints are expressed as inequalities rather than equations. In some problems, not all the design variables may be restricted to be nonnegative. Hence, the first step in solving a linear program is to convert it to a problem in standard form. 4.3.1

Handling Inequalities

Inequality constraints can be converted to equations by introducing slack or surplus variables. For example, an inequality of the type x1 ⫹ 2x2 ⫹ 3x3 ⫹ 4x4 ⭐ 25

(4.1)

can be converted to an equation by introducing a slack variable S1 as follows: x1 ⫹ 2x2 ⫹ 3x3 ⫹ 4x4 ⫹ S1 ⫽ 25 The variable S1 is nonnegative, and it represents the slack between the lefthand side and the right-hand side of inequality (4.1). Similarly, an inequality of the type 2x1 ⫹ x2 ⫺ 3x3 ⭓ 12 can be converted to an equation by introducing a surplus variable S2 as follows: 2x1 ⫹ x2 ⫺ 3x3 ⫺ S2 ⫽ 12 It should be emphasized here that slack and surplus variables are as much a part of the original problem as the design variables used in the formulation of the linear program. These variables can remain positive throughout, and their values in the optimal solution give useful information as to whether or not the inequality constraints are binding or active in the problem. 4.3.2

Handling Unrestricted Variables

In some situations, it may become necessary to introduce a variable that can assume both positive and negative values. Since the standard form requires all variables to be nonnegative, an unrestricted variable is generally replaced by the difference of two nonnegative variables. For example, if s1 is an unrestricted variable, then the following transformation of variables is used for s1:

160

LINEAR PROGRAMMING

s1 ⫽ s1⫹ ⫺ s1⫺ s⫹ 1 ⭓ 0

s1⫺ ⭓ 0

The value of s1 is positive or negative depending on whether s⫹ 1 is larger or smaller than s⫺ . 1 As an illustration, consider the following nonstandard linear program. Example 4.5

Maximize

Z ⫽ x1 ⫺ 2x2 ⫹ 3x3

Subject to

x1 ⫹ x2 ⫹ x3 ⭐ 7

(4.2)

x1 ⫺ x2 ⫹ x3 ⭓ 2

(4.3)

3x1 ⫺ x2 ⫺ 2x3 ⫽ ⫺5

(4.4)

x1, x2 ⭓ 0 where x3 is unrestricted in sign. To convert the above problem to standard form: 1. Replace x3 by x4 ⫺ x5, where x4, x5 ⭓ 0. 2. Multiply both sides of Eq. (4.4) by ⫺1. 3. Introduce slack and surplus variables x6 and x7 to constraints (4.2) and (4.3), respectively. 4. Assign zero profit for x6 and x7 so that the objective function is not altered. Thus, the problem reduces to the following standard linear program: Maximize Subject to

Z ⫽ x1 ⫺ 2x2 ⫹ 3x4 ⫺ 3x5 x1 ⫹ x2 ⫹ x4 ⫺ x5 ⫹ x6

⫽7

x1 ⫺ x2 ⫹ x4 ⫺ x5

⫺ x7 ⫽ 2

⫺3x1 ⫹ x2 ⫹ 2x4 ⫺ 2x5

⫽5

x1, x2, x4, x5, x6, x7 ⱖ 0 Let us review the basic definitions using the standard form of the linear programming problem given by Maximize

Z ⫽ cx

Subject to Ax ⫽ b

x⭓0

4.4

PRINCIPLES OF THE SIMPLEX METHOD

161

1. A feasible solution is a nonnegative vector x satisfying the constraints Ax ⫽ b. 2. A feasible region, denoted by S, is the set of all feasible solutions. Mathematically, S ⫽ {x兩Ax ⫽ b, x ⭓ 0}

3.

4.

5.

6. 7.

4.4

If the feasible set S is empty, then the linear program is said to be infeasible. An optimal solution is a vector xo such that it is feasible, and its value of the objective function (cxo) is larger than that of any other feasible solution. Mathematically, xo is optimal if and only if xo 僆 S, and cxo ⭓ cx for all x 僆 S. The optimal value of a linear program is the value of the objective function corresponding to the optimal solution. If Zo is the optimal value, then Zo ⫽ cxo. When a linear program has more than one optimal solution, it is said to have alternate optimal solutions. In this case, there exists more than one feasible solution having the same optimal value (Zo) of the objective function. The optimal solution of a linear program is said to be unique when no other optimal solution exists. When a linear program does not possess a finite optimum (i.e., max Z → ⫹⬁ or min Z → ⫺⬁), it is said to have an unbounded optimum. PRINCIPLES OF THE SIMPLEX METHOD

Consider the general LP problem in standard form with m equations and n variables: Maximize

Z ⫽ c1x1 ⫹ c2x2 ⫹ ⫹ cn xn

Subject to a11x1 ⫹ a12x2 ⫹ ⫹ a1n xn ⫽ b1 a21x1 ⫹ a22x2 ⫹ ⫹ a2n xn ⫽ b2 ⯗

am1x1 ⫹ am2x2 ⫹ ⫹ amn xn ⫽ bm x1, . . . , xn ⭓ 0 In general, there are more variables than equations (i.e., m ⬍ n), and consequently the problem has an infinite number of feasible solutions. Hence, the selection of the best feasible solution, which maximizes Z, is a nontrivial problem.

162

LINEAR PROGRAMMING

A classical method of generating the solutions to the constraint equations is the Gauss–Jordan elimination scheme (see Appendix C for a review). The basic principle of the Gauss–Jordan elimination scheme is to reduce the system of m equations in n unknowns to a canonical or row echelon form by elementary row operations. One such canonical reduction using the first m variables (x1, x2, . . . , xm) is ⫹ a1,m⫹1xm⫹1 ⫹ ⫹ a1sxs ⫹ ⫹ a1n xn ⫽ b1

x1 xr

⫹ ar,m⫹1xm⫹1 ⫹ ⫹ arsxs ⫹ ⫹ arn xn ⫽ br

(S1)

xm ⫹ am,m⫹1xm⫹1 ⫹ ⫹ amsxs ⫹ ⫹ amn xn ⫽ bm The variables x1, . . . , xm, which appear with a unit coefficient in only one equation and with zeros in all the other equations, are called basic or dependent variables. In the canonical form, there is one such basic variable in every equation. The other n ⫺ m variables (xm⫹1, . . . , xn) are called nonbasic or independent variables. The advantage of obtaining a canonical form is that multiple solutions to the constraint equations can be obtained by simply choosing different values for the independent variables and solving the canonical system for the dependent variables. There are two types of elementary row operations that can be used to obtain a canonical or row echelon form: 1. Multiply any equation in the system by a positive or negative number. 2. Add to any equation a constant multiple (positive or negative) of any other equation in the system. Definition A pivot operation is a sequence of elementary row operations that reduces the coefficient of a specified variable to unity in one of the equations and zero elsewhere. Definition The solution obtained from a canonical system by setting the nonbasic, or independent, variables to zero is called a basic solution. For example, in system (S1) a basic solution is given by x1 ⫽ b1, . . . , xm ⫽ bm, xm⫹1 ⫽ ⫽ xn ⫽ 0. If the values of b1, . . . , bm are also nonnegative, then this basic solution will be called a basic feasible solution.

4.4


163

Definition A basic feasible solution is a basic solution in which the values of the basic, or dependent, variables are nonnegative. In system (S1), a canonical form was obtained using the first m variables as basic variables. The choice of x1, . . . , xm as basic variables was purely for convenience. As a matter of fact, any set of m variables could have been selected as basic variables out of the possible n variables to obtain a canonical system and a basic solution. This means that with m constraints and n variables, the maximum number of basic solutions to the standard linear program is finite and is given by

冉冊

n! n ⫽ m m!(n ⫺ m)!

By definition, every basic feasible solution is also a basic solution. Hence, the maximum number of basic feasible solutions is also limited by the above expression. At the end of Section 4.2 it was pointed out that whenever there is an optimal solution to a linear program, one of the corner points of the feasible region is always optimal. It can be shown that every corner point of the feasible region corresponds to a basic feasible solution of the constraint equations. This means that an optimal solution to a linear program can be obtained by merely examining its basic feasible solutions. This will be a finite process since the number of basic feasible solutions cannot exceed (nm). A naive approach to solving a linear program (which has an optimal solution) would be to generate all possible basic feasible solutions through canonical reduction and determine which one gives the best objective function value. But the simplex method for solving linear programs does this in a more efficient manner by examining only a fraction of the total number of basic feasible solutions! We shall now develop the details of the simplex method. The simplex method as developed by G. B. Dantzig is an iterative procedure for solving LP problems expressed in standard form. In addition to the standard form, the simplex method requires that the constraint equations be expressed in canonical form from which a basic feasible solution could be obtained readily. The general steps of the simplex method are as follows: 1. Start with an initial basic feasible solution in canonical form. 2. Improve the initial solution if possible by finding another basic feasible solution with a better objective function value. At this step the simplex method implicitly eliminates from consideration all those basic feasible solutions whose objective function values are worse than the present

164

LINEAR PROGRAMMING

one. This makes the procedure more efficient than the naive approach mentioned earlier. 3. Continue to find better basic feasible solutions, improving the objective function values. When a particular basic feasible solution cannot be improved further, it becomes an optimal solution, and the simplex method terminates. We begin by considering the details of step 2 of the simplex method by assuming that the basic solution given by system (S1) is feasible. Thus, we have an initial basic feasible solution in canonical form as follows: Basic:

xi ⫽ bi ⭓ 0

for i ⫽ 1, 2, . . . , m

Nonbasic:

xj ⫽ 0

for j ⫽ m ⫹ 1, . . . , n

The set of basic variables will be called a basis and will be denoted xB. Let the objective function coefficients of the basic variables be denoted cB. For the initial basis, xB ⫽ (x1, . . . , xm)

and

cB ⫽ (c1, c2, . . . , cm)

Since the nonbasic variables are zero, the value of the objective function Z corresponding to the initial basic feasible solution is given by Z ⫽ cBxB ⫽ c1b1 ⫹ ⫹ cmbm Given this, the simplex method checks whether it is possible to find a better basic feasible solution with a higher value of Z. This is done by first examining whether the present solution is optimal. In case it is not optimal, the simplex method obtains an adjacent basic feasible solution with a larger value of Z (or at least as large). Definition An adjacent basic feasible solution differs from the present basic feasible solution in exactly one basic variable. To obtain an adjacent basic feasible solution, the simplex method makes one of the basic variables a nonbasic variable and brings in a nonbasic variable as a basic variable in its place. The problem is to select the appropriate basic and nonbasic variables such that an exchange between them will give the maximum improvement to the objective function. In any basic feasible solution, the basic variables can assume positive values while the nonbasic variables are always held at zero. Hence, making a

4.4


165

nonbasic variable a basic variable is equivalent to increasing its value from zero to a positive quantity. Of course, the choice is made based on which nonbasic variable can improve the value of Z. This is determined by increasing the nonbasic variable by one unit and examining the resulting change in the value of the objective function. To illustrate, consider the nonbasic variable xs. Let us increase its value from 0 to 1 and study its effect on the objective function. Since we are interested in examining adjacent basic feasible solutions only, the values of the other nonbasic variables will continue to remain zero and system (S1) can be rewritten as follows: ⫹a1sxs

x1

⫽ b1 ⯗

⫹ arsxs

xr

⫽ br

(S2)

⯗

xm ⫹ amsxs ⫽ bm From system (S2), we obtain the new solution when xs increases from 0 to 1: xi ⫽ bi ⫺ ais

for i ⫽ 1, . . . , m

xs ⫽ 1 xj ⫽ 0

for j ⫽ m ⫹ 1, . . . , n and j ⫽ s

The new value of the objective function is

冘 c (b ⫺ a ) ⫹ c m

Z⫽

i

i

is

s

i⫽1

Hence, the net change in the value of Z per unit increase in xs, denoted cs, is cs ⫽ new value of Z ⫺ old value of Z

冘 c (b ⫺ a ) ⫹ c ⫺ 冘 c b ⫽c ⫺冘ca m

⫽

m

i

i

is

i⫽1

s

i

i

i⫽1

m

s

i

is

(4.5)

i⫽1

where cs is called the relative profit of the nonbasic variable xs, as opposed to its actual profit cs in the objective function. If cs ⬎ 0, then the objective

166

LINEAR PROGRAMMING

function Z can be increased further by making xs a basic variable. Equation (4.5), which gives the formula for calculating the relative profits, is known as the inner product rule. Inner Product Rule The relative profit coefficient of a nonbasic variable xj, denoted by cj, is given by cj ⫽ cj ⫺ cBPj where cB corresponds to the profit coefficients of the basic variables and Pj corresponds to the jth column in the canonical system of the basis under consideration. Condition of Optimality In a maximization problem, a basic feasible solution is optimal if the relative profits of its nonbasic variables are all negative or zero. It is clear that when all cj ⭐ 0 for nonbasic variables, then every adjacent basic feasible solution has an objective function value lower than the present solution. This implies that we have a local maximum at this point. Since the objective function Z is linear, a local maximum automatically becomes the global maximum. For the sake of illustration, let us assume that cs ⫽ max cj ⬎ 0, and hence the initial basic feasible solution is not optimal. The positive relative profit of xs implies that Z will be increased by cs units for every unit increase of the nonbasic variable xs. Naturally, we would like to increase xs as much as possible so as to get the largest increase in the objective function. But as xs increases, the values of the basic variables change, and their new values are obtained from system (S2) as xi ⫽ bi ⫺ aisxs

for i ⫽ 1, . . . , m

(4.6)

If ais ⬍ 0, then xi increases as xs is increased, and if ais ⫽ 0, the value of xi does not change. However, if ais ⬎ 0, xi decreases as xs is increased and may turn negative (making the solution infeasible) if xs is increased indefinitely. Hence, the maximum increase in xs is given by the following rule: max xs ⫽ min ais⬎0

Let

冋册 bi ais

(4.7)

4.4


167

冋册

br bi ⫽ min ars ais⬎0 ais

Hence, when xs is increased to br / ars, the basic variable xr turns zero first and is replaced by xs in the basis. Then xs becomes the new basic variable in row r, and the new basic feasible solution is given by xi ⫽ bi ⫺ ais xs ⫽

冉冊 br ars

for all i

br ars

(4.8) for all other nonbasic variables j ⫽ m ⫹ 1, . . . , n and j ⫽ s

xj ⫽ 0

Since a unit increase in xs increases the objective function Z by cs units, the total increase in Z is given by

冉冊

(cs)

br ⬎0 ars

Equation (4.7), which determines the basic variables to leave the basis, is known as the minimum-ratio rule in the simplex algorithm. Once again, the simplex algorithm checks whether the basic feasible solution given by (4.8) is optimal by calculating the relative profits for all the nonbasic variables, and the cycle is repeated until the optimality conditions are reached. Let us now illustrate the basic steps of the simplex algorithm with an example. Example 4.6

Let us use the simplex method to solve the following problem: Maximize

Z ⫽ 3x1 ⫹ 2x2

Subject to ⫺x1 ⫹ 2x2 ⭐ 4 3x1 ⫹ 2x2 ⭐ 14 x1 ⫺ x2 ⭐ 3 x1 ⭓ 0

x2 ⭓ 0

Converting the problem to standard form by the addition of slack variables, we get

168

LINEAR PROGRAMMING

Maximize

Z ⫽ 3x1 ⫹ 2x2

Subject to ⫺x1 ⫹ 2x2 ⫹ x3 3x1 ⫹ 2x2

⫽4 ⫹ x4

x1 ⫺ x2

⫽ 14 ⫹ x5 ⫽ 3

x1 ⭓ 0

x2 ⭓ 0

x4 ⭓ 0

x5 ⭓ 0

x3 ⭓ 0

We have a basic feasible solution in canonical form, with x3, x4, and x5 as basic variables. The various steps of the simplex method can be carried out in a compact manner by using a tableau form to represent the constraints and the objective function. In addition, the various calculations (e.g., inner product rule, minimum-ratio rule, pivot operation) can be made mechanical. The use of the tableau form has made the simple method more efficient and convenient for computer implementation. The tableau presents the problem in a detached coefficient form. Tableau 1 gives the initial basic feasible solution for Example 4.6. Note that only the coefficients of the variables are written for each constraint. The objective function coefficients cj are written above their respective xj’s. Basis denotes the basic variables of the initial tableau, namely x3 for constraint 1, x4 for constraint 2, and x5 for constraint 3. The parameter cB denotes their respective cj values. From Tableau 1, we obtain the initial basic feasible solution immediately as x3 ⫽ 4, x4 ⫽ 14, x5 ⫽ 3, and x1 ⫽ x2 ⫽ 0. The value of the objective function Z is given by the inner product of the vectors cB and constants:

冢冣

4 Z ⫽ (0, 0, 0) 14 ⫽ 0 3 Tableau 1

cj 3

2

0

0

0

cB

Basis

x1

x2

x3

x4

x5

Constants

0 0 0

x3 x4 x5

⫺1

3 ①

2 2 ⫺1

1 0 0

0 1 0

0 0 1

4 14 3

3

2

0

0

0

Z⫽0

c row

4.4


169

To check whether the above basic feasible solution is optimal, we calculate the relative profits using the inner product rule. Note that the relative profits c3, c4, c5 will be zero since they are the basic variables. Both the nonbasic variables x1 and x2 have positive relative profits. Hence the current basic feasible solution is not optimal. The nonbasic variable x1 gives the largest per-unit increase in Z, and hence we choose it as the new basic variable to enter the basis. To decide which basic variable is going to be replaced, we apply the minimum-ratio rule [Eq. (4.7)] by calculating the ratios for each constraint that has a positive coefficient under the x1 column as follows: Row Number

Basic Variable

Ratio

1 2 3

x3 x4 x5

⬁ 14 –– 3 –31

Note that no ratio is formed for the first row or is set to ⬁, since the coefficient for x1 in row 1 is negative. This means that x1 can be increased to any amount 14 without making x3 negative. On the other hand, the ratio –– 3 for the second 14 row implies that x4 will become zero when x1 increases to –– 3 . Similarly, as x1 increases to 3, x5 turns zero. The minimum ratio is 3, or as x1 is increased from 0 to 3, the basic variable x5 will turn zero first and will be replaced by x1. Row 3 is called the pivot row, and the coefficient of x1 in the pivot row is called the pivot element (circled in Tableau 1). We obtain the new basic variable x3, x4, and x1 and the new canonical system (Tableau 2) by performing a pivot operation as follows: 1. Add the pivot row (row 3) to the first row to eliminate x1. 2. Multiply the pivot row by ⫺3 and add it to the second row to eliminate x1. To check whether the basic feasible solution given by Tableau 2 is optimal, we must calculate the new relative-profit coefficients (c row). We can do this by applying the inner product rule as before. On the other hand, it is also possible to calculate the new c row through the pivot operation. Since x1 is the new basic variable, its relative-profit coefficient in Tableau 2 should be zero. To achieve this, we multiply the third row of Tableau 1 (the pivot row) by ⫺3 and add it to the c row. This will automatically give the new c row for Tableau 2! The movement from Tableau 1 to Tableau 2 is illustrated pictorially in Figure 4.5. The feasible region of Example 4.6 is denoted ABCDE. The dotted lines correspond to the objective function lines at Z ⫽ 3 and Z ⫽ 9. Note that the basic feasible solution represented by Tableau 1 corresponds to the corner

170

LINEAR PROGRAMMING

Tableau 2

cj 3

2

0

0

0

cB

Basis

x1

x2

x3

x4

x5

0 0 3

x3 x4 x1

0 0 1

1 ⑤ ⫺1

1 0 0

0 1 0

⫺3

0

5

0

0

⫺3

c row

1 1

Constants 7 5 3 Z⫽9

point A. The c row in Tableau 1 indicates that either x1 or x2 may be made a basic variable to increase Z. From Figure 4.5 it is clear that either x1 or x2 can be increased to increase Z. Having decided to increase x1, it is clear that we cannot increase x1 beyond three units (point B) in order to remain in the feasible region. This was essentially the minimum-ratio value that was obtained by the simplex method. Thus, Tableau 2 is simply corner point B in Figure 4.5.

Figure 4.5. Feasible region for Example 4.6.

4.4


171

Still the c row has a positive element, indicating that the nonbasic variable x2 can improve the objective function further. To apply the minimum-ratio rule, we find the minimum of (–71 , –55 , ⬁). This implies that x2 replaces the basic variable x4. The next basic feasible solution after the pivot operation is given by Tableau 3. Since none of the coefficients in the c row is positive, this tableau is optimal. An optimal solution is given by x1 ⫽ 4, x2 ⫽ 1, x3 ⫽ 6, x4 ⫽ 0, x5 ⫽ 0, and the optimal value of Z is 14. Alternate Optima. In Tableau 3, the nonbasic variable x5 has a relative profit of zero. This means that no increase in x5 will produce any change in the objective function value. In other words x5 can be made a basic variable and the resulting basic feasible solution will also have Z ⫽ 14. By definition, any feasible solution whose value of Z equals the optimal value is also an optimal solution. Hence, we have an alternate optimal solution to this linear program. This can be obtained by making x5 a basic variable. By the minimum-ratio rule, x3 leaves the basis, and we have an alternative optimal tableau, as shown in Tableau 4. Thus the alternate optimal solution is given by x1 ⫽ –52 , x2 ⫽ 13 15 –– 4 , x3 ⫽ 0, x4 ⫽ 0, and x5 ⫽ –– 4 . In general, an alternate optimal solution is indicated whenever there exists a nonbasic variable whose relative profit (cj coefficient) is zero in the optimal tableau. Referring to Figure 4.5, the movement from Tableau 2 to Tableau 3 by the simplex method corresponds to moving along the edge BC of the feasible region. Using the objective function Z ⫽ 3x1 ⫹ 2x2, it is clear that corner points C and D, as well as the line CD, are optimal. Summary of Computational Steps

In summary, the computational steps of the simplex method in tableau form for a maximization problem are as follows: Step 1. Express the problem in standard form. Step 2. Start with an initial basic feasible solution in canonical form and set up the initial tableau. (We shall discuss later in the section how to find an

Tableau 3

cj 3

2

0

0

0

cB

Basis

x1

x2

x3

x4

x5

Constants

0 2 3

x3 x2 x1

0 0 1

0 1 0

1 0 0

⫺–15 –15 –15

–85 ⫺–35 –25

6 1 4

0

0

0

⫺1

0

c row

Z ⫽ 14

172

LINEAR PROGRAMMING

Tableau 4

cj 3

2

0 x3

cB

Basis

x1

x2

0 2 3

x5 x2 x1

0 0 1 0

c row

0

0

x4

x5

Constants

0 1 0

5 8 3 8 1 4

– – ⫺–

1 8 1 8 1 4

⫺– – –

1 0 0

15 –– 4 13 –– 4 –52

0

0

⫺1

0

Z ⫽ 14

initial basic feasible solution in canonical form if none exists by inspection.) Step 3. Use the inner product rule to find the relative-profit coefficients (c row). Step 4. If all the cj coefficients are nonpositive, the current basic feasible solution is optimal. Otherwise, select the nonbasic variable with the most positive cj value to enter the basis. Step 5. Apply the minimum-ratio rule to determine the basic variable to leave the basis. Step 6. Perform the pivot operation to get the new tableau and the basic feasible solution. Step 7. Compute the relative-profit coefficients by using the pivot operation or the inner product rule. Return to step 4. Each sequence of steps 4–7 is called an iteration of the simplex method. Thus each iteration gives a new tableau and an improved basic feasible solution. The efficiency of the simplex method depends on the number of basic feasible solutions it examines before reaching the optimal solution. Hence, the number of iterations is an important factor in simplex calculations. 4.4.1

Minimization Problems

There are two approaches to solving a minimization problem: Approach 1. Convert the minimization problem to an equivalent maximization problem by multiplying the objective function by ⫺1 and then use the

simplex method as outlined for a maximization problem. Approach 2. Recall that the coefficients in the c row give the net change in

the value of Z per unit increase in the nonbasic variable. A negative coefficient in the c row indicates that the corresponding nonbasic variable (when increased) will reduce the value of the objective function. Hence, in minimi-

4.4


173

zation problems, only those nonbasic variables with negative cj values are eligible to enter the basis and improve the objective function. The optimal solution is obtained when all the coefficients in the c row are nonnegative. Thus all seven steps of the simplex method can be used for solving minimization problems with a minor modification in step 4. Modified Step 4. If all the coefficients in the c row are positive or zero, the

current basic feasible solution is optimal. Otherwise, select the nonbasic variable with the lowest (most negative) value in the c row to enter the basis. 4.4.2

Unbounded Optimum

A linear program is said to have an unbounded optimum when its optimal value is not finite and it tends to ⫹⬁ or ⫺⬁ depending on whether the objective is maximized or minimized, respectively. Existence of an unbounded optimum is recognized by the simplex method when the minimum-ratio rule [given by Eq. (4.7)] fails to identify the basic variable to leave the basis. This happens when none of the constraint coefficients of the nonbasic variable selected to enter the basis is positive. This means that no finite ratios can be formed in Eq. (4.7), and the nonbasic variable selected to improve the objective function can be increased indefinitely without violating the feasibility of the constraint set. This implies that the optimal value of the linear program is not finite. Example 4.7

Consider the simplex tableau of a maximization problem shown in Tableau 5. Since c1 ⬎ 0, the nonbasic variable x1 can enter the basis to increase Z. But the minimum-ratio rule fails, as there is no positive entry under the x1 column. In other words, as x1 increases, both the basic variables x3 and x2 will also increase and hence can never become negative to limit the increase of x1. This means that x1 can be increased indefinitely. Since each unit increase on x1 increases Z by 11 units, Z can be increased indefinitely, and we have an unbounded optimum to the linear program. Note that an unbounded optiTableau 5

cj 2

3

0

0

cB

Basis

x1

x2

x3

x4

0 3

x3 x2

⫺2 ⫺3

0 1

1 0

0 1

11

0

0

⫺3

c row

Constants 6 4 Z ⫽ 12

174

LINEAR PROGRAMMING

mum for a practical problem implies an error in formulation of the linear program, in particular, omission of some important constraints. 4.4.3

Degeneracy and Cycling

A basic feasible solution in which one or more of the basic variables are zero is called a degenerate basic feasible solution. In contrast, a basic feasible solution in which all the basic variables are positive is said to be nondegenerate. Degeneracy may be present in the initial formulation of the problem itself (when one or more right-hand side constants are zero) or it may be introduced during simplex calculations whenever two or more rows tie for the minimum ratio. When the basic feasible solution is degenerate, it is quite possible for the minimum ratio to be zero. This implies that when a basis change is performed, there will not be any real improvement in the value of the objective function. As a matter of fact, it is conceivable for the simplex method to go through a series of iterations with no improvement in Z value. This means that costly simplex calculations will have to be performed in the tableaus without any real return. This naturally reduces the computational efficiency of the simplex method. But a more important question is whether it is possible for the simplex method to go on indefinitely without improving the objective function. In fact, examples have been constructed to show that such a thing is theoretically possible. In such situations the simplex method may get into an infinite loop and will fail to reach the optimal solution. This phenomenon is called classical cycling or simply cycling in the simplex algorithm. Fortunately, such cycling does not happen in practical problems in spite of the fact that many practical problems have degenerate solutions. However, in a small number of cases, practical problems have been observed to fail to converge on a computer. This is mainly due to certain characteristics of the computer software and the way in which the simplex method is programmed. Gass [1] has called this computer cycling. Such problems invariably converge when solved on a different computer system. 4.4.4

Use of Artificial Variables

A major requirement of the simplex method is the availability of an initial basic feasible solution in canonical form. Without it the initial simplex tableau cannot be formed. There is a systematic way of getting a canonical system with a basic feasible solution when none is available by inspection. First the LP problem is converted to standard form such that all the variables are nonnegative, the constraints are equations, and all the right-hand side constants are nonnegative. Then each constraint is examined for the existence of a basic variable. If none is available, a new variable is added to act as the

4.4


175

basic variable in that constraint. In the end, all the constraints will have a basic variable, and by definition we have a canonical system. Since the righthand-side elements are nonnegative, an initial simplex tableau can be formed readily. Of course, the additional variables have no relevance or meaning to the original problem. They are merely added so that we will have a ready canonical system to start the simplex method. Hence, these variables are termed artificial variables as opposed to the real decision variables in the problem. Eventually they will be forced to zero lest they unbalance the equations. To illustrate the use of artificial variables, consider the following LP problem. Example 4.8

Z ⫽ ⫺3x1 ⫹ x2 ⫹ x3

Minimize Subject to

x1 ⫺ 2x2 ⫹ x3 ⭐ 11 ⫺4x1 ⫹ x2 ⫹ 2x3 ⭓ 3 ⫺ x3 ⫽ ⫺1

2x1 x1 ⭓ 0

x2 ⭓ 0

x3 ⭓ 0

First the problem is converted to the standard form as follows: Minimize Subject to

Z ⫽ ⫺3x1 ⫹ x2 ⫹ x3 x1 ⫺ 2x2 ⫹ x3 ⫹ x4 ⫺4x1 ⫹ x2 ⫹ 2x3 ⫺2x1

x1 ⭓ 0

⫽ 11 ⫺ x5 ⫽ 3

(4.10)

⫽1

(4.11)

⫹ x3

x2 ⭓ 0

(4.9)

x3 ⭓ 0

x4 ⭓ 0

x5 ⭓ 0

In Eq. (4.9) the slack variable x4 is a basic variable. Since there are no basic variables in the other equations, we add artificial variables x6 and x7 to Eq. (4.10) and (4.11), respectively. To retain the standard form, x6 and x7 will be restricted to be nonnegative. We now have an ‘‘artificial system’’ given by x1 ⫺ 2x2 ⫹ x3 ⫹ x4 ⫺4x1 ⫹ x2 ⫹ 2x3 ⫺2x1

⫽ 11 ⫺ x5 ⫹ x6

⫹ x3

x1 ⭓ 0

⫽3

x7 ⫽ 1 x7 ⭓ 0

176

LINEAR PROGRAMMING

The artificial system has a basic feasible solution in canonical form given by x1 ⫽ x2 ⫽ x3 ⫽ 0, x4 ⫽ 11, x5 ⫽ 0, x6 ⫽ 3, x7 ⫽ 1. But this is not a feasible solution to the original problem, due to the presence of the artificial variables x6 and x7 at positive values. On the other hand, it is easy to see that any basic feasible solution to the artificial system, in which the artificial variables x6 and x7 are zero, is automatically a basic feasible solution to the original problem. Hence, the objective is to reduce the artificial variables to zero as soon as possible. This can be accomplished by the two-phase simplex method. 4.4.5

Two-Phase Simplex Method

This is an approach to handle the artificial variables whenever they are added. Here the LP problem is solved in two phases. Phase 1. This phase consists of finding an initial basic feasible solution to the original problem. In other words, the removal of the artificial variables is taken up first. For this an artificial objective function is created that is the sum of all the artificial variables. The artificial objective is then minimized using the simplex method. In Example 4.8, the two-phase simplex method would create the following phase 1 linear program with W representing the artificial objective function:

Minimize Subject to

W ⫽ x6 ⫹ x7 x1 ⫺ 2x2 ⫹ x3 ⫹ x4 ⫺4x1 ⫹ x2 ⫹ 2x3 ⫺2x1

⫹ x3

⫽ 11 ⫺ x5 ⫹ x6

⫽3 ⫹ x7 ⫽ 1

x1 ⭓ 0 x7 ⭓ 0 The original objective function Z ⫽ ⫺3x1 ⫹ x2 ⫹ x3 is temporarily set aside during the phase 1 solution. If the minimum value of the artificial problem were zero, then all the artificial variables would have been reduced to zero, and we would have a basic feasible solution to the original problem. (Note: If the sum of nonnegative variables is zero, then each variable must be identically equal to zero.) We then go to phase 2. In case the minimum value of the artificial problem is positive, then at least one of the artificial variables is positive. This means that the original problem without the artificial variables is infeasible, and we terminate. Phase 2. In this phase, the basic feasible solution found at the end of phase 1 is optimized with respect to the original objective function. In other words, the final tableau of Phase 1 becomes the initial tableau for Phase 2 after

4.5

COMPUTER SOLUTION OF LINEAR PROGRAMS

177

changing the objective function. The simplex method is once again applied to determine the optimal solution.*

4.5


Many practical problems formulated as linear programs (LP) run into hundreds of constraints and thousands of decision variables. These invariably have to be solved using a digital computer. In tableau form, the simplex method is an iterative procedure and can be applied mechanically to any problem. Hence, it is ideally suited for computer implementation. 4.5.1

Computer Codes

Evolution of Commercial Packages for LP

The development of the simplex method and the birth of the digital computer occurred at about the same time, in the late 1940s. The simplex method would be of no practical use without the digital computer because of the prodigious amount of computation required to solve all but the smallest problems. Solving large LP models has always been one of the most challenging computational tasks for each new generation of computers. This remains true today, as the meaning of ‘‘large’’ expands with the capabilities of each new computer. It is interesting that the interior point method appeared at about the same time as the widespread availability of vector and parallel computers, since it seems much better suited to these new architectures than the simplex method. PC Software for LP

Up until the mid-1980s almost all industrial LP models were solved on mainframe computers, with software provided by the computer manufacturers or by a very small number of specialty LP software companies. In the late 1980s the situation changed drastically. LP models with a few hundred constraints can now be solved on personal computers (PCs using Intel Pentium chips). This has expanded the use of LP dramatically. There are even LP solvers that can be invoked directly from inside spreadsheet packages. For example, Microsoft Excel and Microsoft Office for Windows contain a general purpose optimizer for solving small linear, integer and nonlinear programming problems. The LP optimizer is based on the simplex algorithm. There are now at least a hundred companies marketing LP software for personal computers. * The reader may refer to Ravindran, Phillips, and Solberg [2, pp. 51–52] for the complete solution of Example 4.8 by the two-phase simplex method.

178

LINEAR PROGRAMMING

They can be found in any issue of OR/MS Today, a publication of the Institute for Operations Research and Management Science. For a 2003 survey of LP software, refer to the article by Fourer [3]. A web version of the survey can be reached via Fourer’s home page at http://iems.nwu.edu/ ⬃4er/. High-End LP Software

As we entered the 1990s, another dramatic change became apparent. The ‘‘workstation’’ class of machines offered mainframe speed at a fraction of the mainframe cost. The result is that virtually all currently used LP models can now be solved on workstations. We now have the capability to solve truly enormous LP models using interior point methods (see Section 4.8.3) on supercomputers, but at the moment users are just beginning to realize this and to think of larger models. For example, an oil company that has been solving single-period production planning models for each refinery can start to piece together a multiperiod, multirefinery model that incorporates distribution as well as production planning. Another sure source of very large LP models is stochastic programming, which attempts to deal with uncertainty in the context of optimization. In 1989 three new high-end LP software systems were introduced that dominate all earlier systems and redefine the state-of-the-art. The first is OSL from IBM. OSL incorporates both the simplex method and the interior point method. The simplex part is a very substantial improvement over IBM’s earlier LP software. The second is CPlex, a simplex code written by Robert Bixby of Rice University and marketed by CPlex Optimization. The third is OB1, an interior point code written by Roy Marsten and David Shanno, and marketed by XMP Software LP Modeling Languages

During the 1980s there was also great progress in the development of computer tools for building LP models. The old mainframe LP systems had ‘‘matrix generator’’ languages that provided some assistance in constructing LP coefficient matrices out of collections of data tables. The focus was on the matrix, however, rather than on the algebraic form of the model. The matrix is a very large, very concrete object that is of little real interest to the human model builder. There are now modeling languages that allow the user to express a model in a very compact algebraic form, with whole classes of constraints and variables defined over index sets. Models with thousands of constraints and variables can be defined in a couple of pages, in a syntax that is very close to standard algebraic notation. The algebraic form of the model is kept separate from the actual data for any particular instance of the model. The computer takes over the responsibility of transforming the abstract form of the model and the specific data into a specific coefficient matrix. This has greatly sim-

4.5


179

plified the building, and even more the changing, of LP models. There are several modeling languages available for PCs. The two high-end products are GAMS (General Algebraic Modeling System) and AMPL (A Mathematical Programming Language). GAMS is marketed by GAMS Development. For a reference on GAMS, see Brooke et al. [4]. AMPL is marketed by AT&T. For a general introduction to modeling languages, see Fourer [5] and for an excellent discussion of AMPL, see Fourer et al. [6]. LP Software on the Internet

For optimization software on the internet [7], users can get a complete list of optimization software available for LP problems at the following web site: http://www.ece.northwestern.edu/otc Maintained by Argonne National Lab and Nortwestern Univeristy, this OTC (Optimization Technology Center) site provides access not only to the NEOS (Network Enabled Optimization System) software guide but also to the other optimization related sites that are continually updated. The NEOS guide on optimization software is based on the text book by More and Wright [8], an excellent resource for those interested in a broad review of the various optimization methods and their computer codes. The book is divided into two parts. Part I has an overview of algorithms for different optimization problems categorized as unconstrained optimization, nonlinear least squares, nonlinear equations, linear programming, quadratic programming, bound-constrained optimization, network optimization and integer programming. Part II includes product descriptions of 75 software packages that implement the algorithms described in Part I. Much of the software described in this book is in the public domain and can be obtained through the internet. 4.5.2

Computational Efficiency of the Simplex Method

The computational efficiency of the simplex method depends on (1) the number of iterations (basic feasible solutions) required to reach the optimal solution, and (2) the total computer time needed to solve the problem. Much effort has been spent on studying computational efficiency with regard to the number of constraints and the decision variables in the problem. Empirical experience with thousands of practical problems shows that the number of iterations of a standard linear program with m constraints and n variables varies between m and 3m, the average being 2m. A practical upper bound for the number of iterations is 2(m ⫹ n), although occasionally problems have violated this bound. If every decision variable had a nonzero coefficient in every constraint, then the computational time would increase approximately in relation to the cube of the number of constraints, m3. However, large problems have fewer

180

LINEAR PROGRAMMING

than 1% of the matrix coefficients nonzero. The use of sparse matrix techniques makes computation times unpredictable, but far better than m3 would suggest. The computational efficiency of the simplex method is more sensitive to the number of constraints than to the number of variables. Hence the general recommendation is to keep the number of constraints as small as possible by avoiding unnecessary or redundant constraints in the formulation of the LP problem. Unlike the simplex method which moved along the boundary of the feasible region, an interior point LP algorithm was developed in 1979 by the Russian mathematician Khachian [9]. Even though Khachian’s algorithm was theoretically an important breakthrough, it did not turn out to be promising upon computer implementation due to its slow convergence. In 1984, Narendra Karmarkar [10], an AT&T researcher, developed an interior point algorithm, which was claimed to be 50 times faster than the simplex method. By 1990, Karmarkar’s seminal work had spawned hundreds of research papers and a large class of interior point methods. It has become clear that while the initial claims were somewhat exaggerated, interior point methods do become competitive for very large problems. However, since 1984 there have been dramatic improvements in the computer implementations of the simplex method as well, largely spurred by the competition between the two methods. Recent computational studies suggest that the simplex method is more efficient for LP models with a few hundred constraints. Beyond a few thousand constraints, however, the interior point methods become more efficient. Details of the interior point methods are discussed in Section 4.8.3. 4.6

SENSITIVITY ANALYSIS IN LINEAR PROGRAMMING

In all LP models the coefficients of the objective function and the constraints are supplied as input data or as parameters to the model. The optimal solution obtained by the simplex method is based on the values of these coefficients. In practice the values of these coefficients are seldom known with absolute certainty, because many of them are functions of some uncontrollable parameters. For instance, future demands, the cost of raw materials, or the cost of energy resources cannot be predicted with complete accuracy before the problem is solved. Hence the solution of a practical problem is not complete with the mere determination of the optimal solution. Each variation in the values of the data coefficients changes the LP problem, which may in turn affect the optimal solution found earlier. In order to develop an overall strategy to meet the various contingencies, one has to study how the optimal solution will change with changes in the input (data) coefficients. This is known as sensitivity analysis or post-optimality analysis. Other reasons for performing a sensitivity analysis are:

4.6

SENSITIVITY ANALYSIS IN LINEAR PROGRAMMING

181

1. Some data coefficients or parameters of the linear program may be controllable; for example, availability of capital, raw material, or machine capacities. Sensitivity analysis enables us to study the effect of changes in these parameters on the optimal solution. If it turns out that the optimal value (profit/cost) changes (in our favor) by a considerable amount for a small change in the given parameters, then it may be worthwhile to implement some of these changes. For example, if increasing the availability of labor by allowing overtime contributes to a greater increase in the maximum return, as compared to the increased cost of overtime labor, then we might want to allow overtime production. 2. In many cases, the values of the data coefficients are obtained by carrying out statistical estimation procedures on past figures, as in the case of sales forecasts, price estimates, and cost data. These estimates, in general, may not be very accurate. If we can identify which of the parameters affect the objective value most, then we can obtain better estimates of these parameters. This will increase the reliability of our model and the solution. We shall now illustrate the practical uses of sensitivity analysis with the help of an example. Example 4.9

A factory manufactures three products, which require three resources—labor, materials, and administration. The unit profits on these products are $10, $6, and $4, respectively. There are 100 hr of labor, 600 lb of material, and 300 hr of administration available per day. In order to determine the optimal product mix, the following LP model is formulated and solved: Maximize Subject to

Z ⫽ 10x1 ⫹ 6x2 ⫹ 4x3 x1 ⫹ x2 ⫹ x3 ⭐ 100 10x1 ⫹ 4x2 ⫹ 5x3 ⭐ 600 2x1 ⫹ 2x2 ⫹ 6x3 ⭐ 300

(labor) (material) (administration)

x1, x2, x3 ⭓ 0 where x1, x2, and x3 are the daily production levels of products 1, 2, and 3 respectively. A computer output of the solution of Example 4.9 is given in Table 4.1. (For example, a popular LP software for PCs marketed by LINDO Systems [11] will provide an output containing the information in Table 4.1. Micro-

182

LINEAR PROGRAMMING

Table 4.1

x1 ⫽ 33.33, x2 ⫽ 66.67, x3 ⫽ 0 maximum profit ⫽ $733.33 For row 1 ⫽ $3.33, for row 2 ⫽ 0.67, for row 3 ⫽ 0 For x1 ⫽ 0, for x2 ⫽ 0, for x3 ⫽ 2.67

Optimal Solution: Optimal Value: Shadow Prices: Opportunity Costs:

Ranges on Objective Function Coefficients Variable

Lower Limit

Present Value

Upper Limit

x1 x2 x3

6 4

10 6 4

15 10 6.67

⫺⬁

Ranges on Right-Hand-Side (RHS) Constants Row

Lower Limit

Present Value

Upper Limit

1 2 3

60 400 200

100 600 300

150 1000 ⬁

soft’s Excel Solver for LPs can also a provide a similar output.) We note from the optimal solution that the optimal product mix is to produce products 1 and 2 only at levels 33.33 and 66.67 units, respectively. The shadow prices give the net impact in the maximum profit if additional units of certain resources can be obtained. Labor has the maximum impact, providing $3.33 increase in profit for each additional hour of labor. Of course, the shadow prices on the resources apply as long as their variations stay within the prescribed ranges on RHS constants given in Table 4.1. In other words, a $3.33/hr increase in profit is achievable as long as the labor hours are not increased beyond 150 hr. Suppose it is possible to increase the labor hours by 25% by scheduling overtime that incurs an additional labor cost of $50. To see whether it is profitable to schedule overtime, we first determine the net increase in maximum profit due to 25 hr of overtime as (25)($3.33) ⫽ $83.25. Since this is more than the total cost of overtime, it is economical to schedule overtime. It is important to note that when any of the RHS constants is changed, the optimal solution will change. However, the optimal product mix will be unaffected as long as the RHS constant varies within the specified range. In other words, we will still be making products 1 and 2 only, but their quantities may change. The ranges on the objective function coefficients given in Table 4.1 exhibit the sensitivity of the optimal solution with respect to changes in the unit profits of the three products. It shows that the optimal solution will not be affected as long as the unit profit of product 1 stays between $6 and $15. Of course, the maximum profit will be affected by the change. For example, if

4.8

ADDITIONAL TOPICS IN LINEAR PROGRAMMING

183

the unit profit on product 1 increases from $10 to $12, the optimal solution will be the same but the maximum profit will increase to $733.33 ⫹ (12 ⫺ 10)(33.33) ⫽ $799.99. Note that product 3 is not economical to produce. Its opportunity cost measures the negative impact of producing product 3 to the maximum profit. Hence, a further decrease in its profit contribution will not have any impact on the optimal solution or maximum profit. Also, the unit profit on product 3 must increase to $6.67 (present value ⫹ opportunity cost) before it becomes economical to produce.

4.7

APPLICATIONS

Linear programming models are used widely to solve a variety of military, economic, industrial, and social problems. The oil companies are among the foremost users of very large LP models, using them in petroleum refining, distribution, and transportation. The number of LP applications has grown so much in the last 20 years that it will be impossible to survey all the different applications. Instead, the reader is referred to two excellent textbooks, Gass [12] and Salkin and Saha [13], which are devoted solely to LP applications in diverse areas like defense, industry, retail business, agriculture, education, and environment. Many of the applications also contain a discussion of the experiences in using the LP model in practice. An excellent bibliography on LP applications is available in Gass [14], which contains a list of references arranged by area (e.g., agriculture, industry, military, production, transportation). In the area of industrial application the references have been further categorized by industry (e.g., chemical, coal, airline, iron and steel, paper, petroleum, railroad). For another bibliography on LP applications, the readers may refer to a survey by Gray and CullinanJames [15]. For more recent applications of LP in practice, the reader should check the recent issues of Interfaces, AIIE Transactions, Decision Sciences, European Journal of Operational Research, Management Science, Operations Research, Operational Research (U.K.), Naval Research Logistics Quarterly, and OpSearch (India).

4.8


In this section, we briefly discuss some of the advanced topics in linear programming. These include duality theory, the dual simplex method, integer programming and interior point methods. For a full discussion, see references 2, 16, or 17.

184

LINEAR PROGRAMMING

4.8.1

Duality Theory

From both the theoretical and practical points of view, the theory of duality is one of the most important and interesting concepts in linear programming. The basic idea behind the duality theory is that every linear programming problem has an associated linear program called its dual such that a solution to the original linear program also gives a solution to its dual. Thus, whenever a linear program is solved by the simplex method, we are actually getting solutions for two linear programming problems! We shall introduce the concept of a dual with the following linear program. Example 4.10

Maximize Subject to

Z ⫽ x1 ⫹ 2x2 ⫹ 3x3 ⫹ 4x4 x1 ⫹ 2x2 ⫹ 2x3 ⫹ 3x4 ⬉ 20 2x1 ⫹ x2 ⫹ 3x3 ⫹ 2x4 ⬉ 20 x1, . . . , x4 ⭌ 0

The above linear program has two constraints and four variables. The dual of this problem is written as follows: Minimize Subject to

W ⫽ 20y1 ⫹ 20y2 y1 ⫹ 2y2 ⭌ 1 2y1 ⫹ y2 ⭌ 2 2y1 ⫹ 3y2 ⭌ 3 3y1 ⫹ 2y2 ⭌ 4 y1, y2 ⭌ 0

y1 and y2 are called the dual variables. The original problem is called the primal problem. Comparing the primal and the dual problems, we observe the following relationships: 1. The objective function coefficients of the primal problem have become the right-hand-side constants of the dual. Similarly, the right-hand-side constants of the primal have become the cost coefficients of the dual. 2. The inequalities have been reversed in the constraints. 3. The objective is changed from maximization in primal to minimization in dual.

4.8


185

4. Each column in the primal corresponds to a constraint (row) in the dual. Thus the number of dual constraints is equal to the number of primal variables. 5. Each constraint (row) in the primal corresponds to a column in the dual. Hence there is one dual variable for every primal constraint. 6. The dual of the dual is the primal problem. In both of the primal and the dual problems, the variables are nonnegative and the constraints are inequalities. Such problems are called symmetric dual linear programs. Definition A linear program is said to be in symmetric form, if all the variables are restricted to be nonnegative, and all the constraints are inequalities (in a maximization problem the inequalities must be in ‘‘less than or equal to’’ form; while in a minimization problem they must be ‘‘greater than or equal to’’). Symmetric Dual Linear Programs. We shall now give the general representation of the primal-dual problems in symmetric form. Primal Maximize

Z ⫽ cx

Subject to Ax ⬉ b x⭌0 Dual Minimize

W ⫽ yb

Subject to yA ⭌ c y⭌0 where A is an (m ⫻ n) matrix, b is an (m ⫻ 1) column vector, c is a (1 ⫻ n) row vector, x is an (n ⫻ 1) column vector, and y is an (1 ⫻ m) row vector. The general rules for writing the dual of a linear program in symmetric form are summarized below: 1. Define one (nonnegative) dual variable for each primal constraint. 2. Make the cost vector of the primal the right-hand-side constants of the dual. 3. Make the right-hand-side vector of the primal the cost vector of the dual.

186

LINEAR PROGRAMMING

4. The transpose of the coefficient matrix of the primal becomes the constraint matrix of the dual. 5. Reverse the direction of the constraint inequalities. 6. Reverse the optimization direction, that is, change minimizing to maximizing and vice versa. There are a number of important relationships between the solution of the primal problem and its dual. They are useful in investigating the general properties of the optimal solution to a linear program and in testing whether a feasible solution is optimal. Now, we shall state some of the duality theorems that give important relationships between the primal and dual solutions. For a proof of the duality theorems, refer to the text by Ravindran, Phillips, and Solberg [2]. Theorem 4.1

Weak Duality Theorem

Consider the symmetric primal-dual linear programs, max Z ⫽ cx, Ax ⬉ b, x ⭌ 0, and min W ⫽ yb, yA ⭌ c, y ⭌ 0. The value of the objective function of the minimum problem (dual) for any feasible solution is always greater than or equal to that of the maximum problem (primal). From the weak duality theorem we can infer the following important results: Corollary 1 The value of the objective function of the maximum (primal) problem for any (primal) feasible solution is a lower bound to the minimum value of the dual objective. Corollary 2 Similarly the objective function value of the minimum problem (dual) for any (dual) feasible solution is an upper bound to the maximum value of the primal objective. Corollary 3 If the primal problem is feasible and its objective is unbounded (ie., max Z → ⫹⬁), then the dual problem cannot have a feasible solution. Corollary 4 Similarly, if the dual problem is feasible, and is unbounded (i.e., min W →

⫺⬁), then the primal problem is infeasible.

4.8


187

To illustrate the weak duality theorem, consider the primal-dual pair given in Example 4.10: x 01 ⫽ x 02 ⫽ x 03 ⫽ x 04 ⫽ 1 is feasible for the primal and y 01 ⫽ y 02 ⫽ 1 is feasible for the dual. The value of the primal objective is Z ⫽ cx0 ⫽ 10 The value of the dual objective is W ⫽ y0b ⫽ 40 Note that cx0 ⬍ y0b which satisfies the weak duality theorem. Using Corollary 1, the minimum value of W for the dual objective function cannot go below 10. Similarly, from Corollary 2, the maximum value of Z for the primal problem cannot exceed 40. The converse of the results of Corollaries 3 and 4 are also true. Corollary 5 If the primal problem is feasible, and the dual is infeasible, then the primal is unbounded. Corollary 6 If the dual problem is feasible and the primal is infeasible, then the dual is unbounded. Theorem 4.2

Optimality Criterion Theorem

If there exist feasible solutions x0 and y0 for the symmetric dual linear programs such that the corresponding values of their objective functions are equal, then these feasible solutions are in fact optimal solutions to their respective problems. Illustration

Consider the primal-dual pair given in Example 4.10, x 01 ⫽ 0, x 02 ⫽ 0, x 03 ⫽ 4, and x 04 ⫽ 4 is a feasible solution to the primal, while y 01 ⫽ 1.2, and y 02 ⫽ 0.2 is feasible for the dual. The value of Z for the primal ⫽ 28 ⫽ the value of the W for the dual. Hence by Theorem 2, the above feasible solutions are optimal solutions to the primal and the dual problems, respectively. Theorem 4.3

Main Duality Theorem

If both the primal and the dual problems are feasible, then they both have optimal solutions such that their optimal values of the objective functions are equal.

188

LINEAR PROGRAMMING

Theorem 4.4

Complementary Slackness Theorem

Consider the symmetric dual problems in matrix form: Primal Maximize

Z ⫽ cx

Subject to Ax ⬉ b x⭌0 Dual Minimize

W ⫽ yb

Subject to yA ⭌ c y⭌0 where A is an (m ⫻ n) matrix, b(m⫻1) and x(n⫻1) are column vectors, and c(1⫻n) and y(1⫻m) are row vectors. Let x0 and y0 be feasible for the primal and the dual problems respectively. Then x0 and y0 are optimal to their respective problems if and only if (y0A ⫺ c)x0 ⫹ y0(b ⫺ Ax0) ⫽ 0 We shall leave it as an exercise to the reader to verify that the optimal solutions to the primal-dual pair given in Example 4.10 satisfies the complementary slackness theorem. Note

The Kuhn–Tucker optimality conditions, discussed in Chapter 5, are direct extensions of the complementary slackness theorem for nonlinear programming problems. 4.8.2

The Dual Simplex Method

The dual simplex method is a modified version of the simplex method using duality theory. There are instances when the dual simplex method has an obvious advantage over the regular simplex method. It plays an important part in sensitivity analysis, parametric programming, solution of integer programming problems, many of the variants of the simplex method, and solution of some nonlinear programming problems. Details of the dual simplex method are discussed in reference 2.

4.8

4.8.3


189

Interior Point Methods

Interior Point Methods for Linear Programming

Dantzig’s simplex method and its variants, start with an extreme point solution in the feasible region and at each iteration, move to an adjacent extreme point with an improved objective value, always traversing along the boundary of the feasible region. Figure 4.6 illustrates the sequence of points after each iteration (shown as ‘‘dots’’) by the simplex method, where the arrows along the boundary show the path. Examining only adjacent solutions reduces the computational effort at each iteration, since it is limited by the number of edges leading to adjacent vertices. However, the number of iterations can be very large because, for an LP in standard form, the feasible region could have up to C(n, m) ⫽ n!/m! n ⫺ m! extreme points. Unlike the simplex method, the interior point methods, popularly known as Karmarkar’s algorithm, traverse the interior of the feasible region as shown in Fig. 4.6. The sequence of points after each iteration is shown as ‘‘circles’’. The major problem with interior point methods is the need to identify the best among all feasible directions at a given solution. The basic idea is to

Figure 4.6. Difference between exterior and interior point methods.

190

LINEAR PROGRAMMING

reduce the number of iterations even though each iteration may involve huge computational effort. In 1984, Narendra Karmarkar (10), an AT&T researcher, developed an interior point algorithm, called the Projective Scaling algorithm. At the time of introduction, it was claimed that Karmarkar’s algorithm was 50 times faster than the simplex method! By 1990, Karmarkar’s seminal work had spawned hundreds of papers and a large class of interior point methods. It has become clear that while the initial claims were somewhat exaggerated, interior point methods do become competitive for very ‘‘large’’ problems and for certain special classes of problems that have always been particularly difficult for the simplex method. Two such classes are highly degenerate problems (many different algebraic basic feasible solutions correspond to the same geometric corner point) and multiperiod problems that involve decisions in successive time periods that are linked together by inventory transfers. For both these classes, the number of iterations taken by simplex can far exceed the 2m ruleof-thumb. As of this writing, the definition of the class of ‘‘very large’’ problems for which interior point methods dominate is changing almost daily as computer implementations of the interior point method become more and more efficient. However, since 1984, there have been dramatic improvements in the computer implementation of the simplex method as well, largely spurred by the competition between the two methods. As a result, simplex codes are more efficient for LP models with a few hundred constraints. Beyond a few thousand constraints, however, the interior point methods become more competitive. Empirical studies by Marsten et al. [18] suggest that when m ⫹ n ⬍ 2000 (number of constraints plus variables), the simplex method should be preferred. When m ⫹ n ⬎ 10,000, interior point methods should be preferred. However, Bixby [19], based on his studies, claims that there is no single dominant method for very large LP problems. Karmarkar’s Algorithm and Its Variants

Karmarkar’s projective scaling algorithm starts with an interior solution, transforms the feasible region such that the current solution is at the center of the transformed feasible region, and then moves in the direction of steepest descent with a step length that stops short of the boundary of the transformed feasible region. Thereafter, an inverse transformation is applied to map the improved solution back to the original feasible region. The process is repeated until an optimum is obtained with the desired precision. Karmarkar’s algorithm is designed for LP problems expressed in a nonstandard form (min cTx subject to Ax ⫽ 0, 兺 xj ⫽ 1, x ⭓ 0, and the optimal objective function value is zero) and involves a complex projective transformation at each iteration. Early implementation revealed difficulties in converting a standard LP problem to Karmarkar’s structure, which made the

4.8


191

matrix A twice as large. An important class of interior point algorithms, called affine scaling algorithms—variants of Karmarkar’s algorithm—address these difficulties. Primal affine scaling algorithms solve the primal LP problems in standard form, whereas dual affine algorithms solve the dual LP problems in inequality form. Both these algorithms use the affine transformation, which is simpler than the projective transformation. Current research suggests that projective scaling algorithms do not compete favorably, in practical implementation, with affine scaling algorithms and their variants. In this section, we shall discuss the basic principles of interior point methods for LPs, in particular the primal affine scaling algorithm. Primal Affine Scaling Algorithm Consider the LP problem in standard form: Minimize

Z ⫽ cTx

Subject to Ax ⫽ b

(4.12)

x⭓0

(4.13)

where A is an m ⫻ n matrix of full-row rank and b, c, and x are column vectors. Let the feasible region be defined as S ⫽ {x 兩 Ax ⫽ b, x ⭓ 0} Definition x0 is called an interior point of S if x0 belongs to S and every component of x is positive, that is, xj ⬎ 0 for all j ⫽ 1, 2, . . . , n. Just like the simplex method, an interior point algorithm has the following three steps: Step 1. Find an interior point solution to begin the method. Step 2. Generate the next interior point that will be feasible and in the interior, with a lower objective function value. Step 3. Test the new point for optimality. If it is not optimal, repeat step 2 until optimality is reached. We shall first discuss the most important step, namely step 2, of the algorithm. Karmarkar, in the development of his algorithm, provides two important principles that are fundamental to all interior point algorithms.

192

LINEAR PROGRAMMING

1. Centering. It is always advantageous to select an interior point at the ‘‘center’’ of the feasible region, for example, an interior point that is equidistant from the ‘‘walls’’ of the feasible region or the coordinate axes. 2. Steepest Descent Direction. Given that the current point is centered, the best direction to use to decrease the objective function is the direction of steepest descent. Mathematical Background. To explain Karmarkar’s principles further, we need to introduce the concept of direction, moving along a direction, descent direction, and feasible direction.

Definition A vector d is called direction if not all its components are zero. Given a point x0 and a direction d, moving along a direction is equivalent to generating points given by x ⫽ x0 ⫹ ␭d

(4.14)

where ␭ is a positive scalar. This is illustrated in Figure 4.7. Moving along the direction d from x0 is essentially moving parallel to the direction vector as given by the following equation:

Figure 4.7. Moving along a direction.

4.8


193

xT ⫽ (2, 1) ⫹ ␭(1, 1) ⫽ [(2 ⫹ ␭), (4 ⫹ ␭)] For ␭ ⬎ 0, we move along the direction d, and for ␭ ⬍ 0, we move against the direction d. Definition A direction d is a descent direction if moving along that direction decreases the value of the objective function Z. For d to be a descent direction, we must have cTx ⬍ cTx0 Using Eq. (4.14), this implies that cT(x0 ⫹ ␭d) ⬍ cTx0

or

␭cTd ⬍ 0

Since ␭ ⬎ 0, this implies that we have to have cTd ⬍ 0 for d to be a descent direction. If c is not a zero vector, then one can choose d ⫽ ⫺c, in which case the direction vector is a decreasing direction since cTd ⫽ cT(⫺c) ⫽ ⫺c2 ⬍ 0 Note: (i) A nonzero vector that inner products with itself will always be positive. (ii) c is called the length of the vector. Thus, the negative of the objective vector is always a descent direction. Moreover, it can be shown that it is in fact the steepest descent direction among all descent directions. If there are no explicit constraints given by Ax ⫽ b, then one can move in any direction from an interior point by a small amount without violating the nonnegativity constraints. However, we have to make sure that the constraints Ax ⫽ b are satisfied as we move along the direction d. Definition A direction d is a feasible direction if moving along that direction satisfies the explicit LP constraints, Ax ⫽ b, that is, A(x0 ⫹ ␭d) ⫽ b

194

LINEAR PROGRAMMING

Since Ax0 ⫽ b (x0 was a feasible solution to begin with), this implies that ␭ Ad ⫽ 0

or

Ad ⫽ 0

In other words, for d to be a feasible direction, it must satisfy Ad ⫽ 0. Of course, it also has to satisfy x0 ⫹ ␭d ⭓ 0 to satisfy nonnegativity restrictions. However, we will treat them implicitly using a ratio test similar to the minimum-ratio rule in the simplex method. To sum up, given an interior point x0: (i) A feasible direction is a vector d ⫽ 0 that satisfies Ad ⫽ 0. The set of all d that satisfies Ad ⫽ 0 is called the null space of matrix A. (ii) A descent direction is a vector d ⫽ 0 that satisfies cTd ⬍ 0. The steepest descent direction is given by d ⫽ ⫺c. Illustration of Karmarkar’s Principles Example 4.11

Minimize

Z ⫽ ⫺2x1 ⫹ x2

Subject to x1 ⫹ x2 ⭐ 10 x2 ⭐ 5 x1, x2 ⭓ 0 The feasible region of this example is shown in Figure 4.8. Consider the two interior points A and B. It is clear that point A is more ‘‘centered’’ in the feasible region than point B. Using the steepest descent direction, namely, d ⫽ ⫺c ⫽ (2, ⫺1)T, we move to the right perpendicular to the objective function line shown as dashed lines in Fig. 4.8. It is clear that Z decreases substantially by moving along the steepest descent direction from point A. On the other hand, the off-center point B will take us toward the wall of the feasible region before much improvement can be made in Z. How to Center an Interior Point? By a simple ‘‘variable scaling’’ (transformation of variables), we can make an ‘‘off-center’’ point to be ‘‘equidistant’’ from the coordinate axes in a transformed feasible region. To center the point B in Example 4.11, we do the following: Define y1 ⫽ x1 /4 and y2 ⫽ x2. By substituting x1 ⫽ 4y1 and x2 ⫽ y2, we get the scaled LP problem as

4.8


195

Figure 4.8. Feasible region of Example 4.11.

Minimize

Z ⫽ ⫺8y1 ⫹ y2

Subject to

4y1 ⫹ y2 ⭐ 10 y2 ⭐ 5 y1, y2 ⭓ 0

The transformed feasible region is shown in Figure 4.9. Point B(4, 1) in the original feasible region is now point (1, 1) in the new region. It is at unit distance from the coordinate axes and is now more centered in the new feasible region. A steepest descent direction move will produce a new point with a substantial decrease in Z. The new point will then be retransformed back to the original feasible region. Scaling the Standard Linear Program

Consider the standard LP problem given by Eqs. (4.12) and (4.13). Let x0 be the initial interior point. To center x0, define

196

LINEAR PROGRAMMING

Figure 4.9. Feasible region of scaled linear program.

x ⫽ Dy where

冤

x01

D⫽

x02

x0n

冥

Here D is a diagonal matrix, called the scaling matrix, whose diagonal elements are the values of the variables at the current interior point x0. Note that

冤

1/x01

D ⫽ DT

and

D⫺1 ⫽

1/x20

1/x0n

冥

4.8


197

Hence, y0 ⫽ D⫺1 x0 ⫽ (1, 1, 1, . . . , 1)T is at unit distance from each coordinate axis. The above transformation is called an affine scaling transformation, from where the interior point algorithm gets its name. The scaled LP problem in variable y is given by Z ⫽ pTy

Minimize

Subject to By ⫽ b

(4.15) y⭓0

(4.16)

where B ⫽ AD and p ⫽ Dc. The original interior point x0 has become y0, a vector of 1’s. To improve 0 y with respect to Z, we need (i) a feasible direction and (ii) a descent direction. To get a feasible direction from y0, a projection matrix P, as defined below, is used: P ⫽ In ⫺ BT(BBT)⫺1B

(4.17)

The projection matrix P satisfies PT ⫽ P and P2 ⫽ PP ⫽ P. Consider a projection vector d obtained using the projection matrix: d ⫽ Pz

(4.18)

where z is any direction vector. We shall show that d will be a feasible direction to the scaled LP problem. To be a feasible direction, d has to satisfy the condition Bd ⫽ 0: Bd ⫽ BPz ⫽ B[In ⫺ BT(BBT)⫺1B]z ⫽ [B ⫺ {(BBT)(BBT)⫺1}B]z ⫽ [B ⫺ B]z ⫽ 0

Since Eq. (4.18) applies to any direction vector z, we can make d a steepest descent direction by choosing z ⫽ ⫺p. For d ⫽ ⫺Pp to be a descent direction, we have to prove that it satisfies the condition pTd ⬍ 0: pTd ⫽ pT(⫺Pp) ⫽ ⫺pTPp ⫽ ⫺pTP2p since P2 ⫽ P ⫽ ⫺(pP)T(Pp) ⫽ ⫺Pp2 ⬍ 0

Thus, the direction vector given by d0 ⫽ ⫺Pp

(4.19)

is a feasible direction and a descent direction as well. Thus, the new points generated by d0 will be given by the equation

198

LINEAR PROGRAMMING

y ⫽ y0 ⫹ ␭d0

for ␭ ⬎ 0

(4.20)

and they will satisfy By ⫽ b and pTy ⬍ pTy0. However, we have to choose ␭ such that we stay inside the feasible region, that is, satisfy y ⭓ 0. We do this by using a ratio test similar to the minimumratio rule used in the simplex method. Using Eq. (4.20), we can find the maximum ␭ such that the solution stays feasible: y ⫽ y0 ⫹ ␭d0 ⭓ 0 or yj ⫽ yj0 ⫹ ␭d j0 ⭓ 0 Since y0 is feasible, y0j ⭓ 0 for all j. If d 0j ⭓ 0, then yj ⭓ 0 since ␭ ⬎ 0. However, if d 0j ⬍ 0, then ␭ must satisfy

␭⭐

y 0j ⫺d 0j

Thus,

␭max ⫽ min d 0j⬍0

y0j ⫺d 0j

If all elements of the vector d0 are nonnegative, then the LP problem is unbounded (recall the analogy in the simplex method when the minimumratio rule fails!). If we chose ␭ ⫽ ␭max, then one or more variables will be zero, and the new point will no longer be an interior point. To avoid that, we set ␭ at 90–95 percent of the maximum value given by the ratio test. This guarantees that the new point y1 ⬎ 0, and we will have a new interior point. Note: The bulk of the computational effort in interior point algorithms is spent in computing the P matrix given by Eq. (4.17). Example 4.12

Consider the LP problem given in Example 4.11. Rewriting it in standard form we, get

4.8


199

Z ⫽ ⫺2x1 ⫹ x2

Minimize

Subject to x1 ⫹ x2 ⫹ x3 x2 ⫹

⫽ 10

x4 ⫽ 5

x1, x2, x3, x4 ⭓ 0 Here, A⫽ b⫽

冋冉冊

册冉冊

1 1 1 0 0 1 0 1 10 5

c⫽

⫺2

1

x ⫽ [x1, x2, x3, x4]T Iteration 1. Let the initial interior point be given by x0 ⫽ (4, 1, 5, 4)T and Z ⫽ cTx0 ⫽ ⫺7. Then

冤

4 0 0 0 0 1 0 0 D⫽ 0 0 5 0 0 0 0 4

冥

where D is a 4 ⫻ 4 scaling matrix. The scaled LP problem is given by Z ⫽ pTy

Minimize

Subject to By ⫽ b

y⭓0

where

册冤

冥冤冥冤冥冤冥冤冥冤冥冤冥冋

⫺2

4 0 0 0 0 1 0 0 p ⫽ Dc ⫽ 0 0 5 0 0 0 0 4

1 ⫽ 0 0

–14

y0 ⫽ D⫺1x0 ⫽

冋

册

4 0 0 0 0 1 0 0 4 1 5 0 ⫽ 0 0 5 0 0 1 0 4 0 0 0 4

1 1 1 0 B ⫽ AD ⫽ 0 1 0 1

1

–15

–14

⫺8

1 0 0

4 1 1 1 ⫽ 5 1 4 1

Z0 ⫽ pTy0 ⫽ ⫺7 To determine a feasible direction, compute the projection matrix P:

200

LINEAR PROGRAMMING

P ⫽ I4 ⫺ BT(BBT)⫺1B

冉

42 1

BBT ⫽

1 17

冤冤

冊

冥冋

4 0 1 1 1 B (BB ) B ⫽ 713 5 0 0 4 T

T ⫺1

1 ⫽ 713

[BBT]⫺1 ⫽

and

册冋

17 ⫺1 42

⫺1

272 64 64 57 340 80 ⫺16 164

340 80 425 ⫺20

冉册

冊

1 17 ⫺1 713 ⫺1 42

4 1 5 0 0 1 0 4

⫺16

164 ⫺20

冥

672

Hence,

P⫽

冤

冥

0.0224 0.6185 ⫺0.0898 ⫺0.4769 0.92 ⫺0.1122 ⫺0.23 ⫺0.1122 0.4039 0.0281 0.0224 ⫺0.23 0.0281 0.0575

⫺0.0898 ⫺0.4769

Verify that BP ⫽ 0. Hence, a direction vector d ⫽ Pz for any z ⫽ 0 will satisfy Bd ⫽ 0 and will be a feasible direction. Compute the projected gradient vector as

冤冥冤冥 ⫺8

d0 ⫽ ⫺Pp ⫽ ⫺P

5.0378 ⫺1.6384 1 ⫽ 0 ⫺3.703 0 0.4092

Note that pTd0 ⫽ ⫺41.92 ⬍ 0. Hence, d0 is not only feasible but also a descent direction. Compute the new feasible point y1 as follows: y1 ⫽ y0 ⫹ ␭d0 ⭓ 0

冢冣冢冣

1 5.0378 1 ⫺1.6384 ⫽ ⫹␭ ⭓0 1 ⫺3.703 1 0.4092 Maximum ␭ ⫽ min (⬁, 1/1.6384, 1/3.703, ⬁) ⫽ 0.27. Set ␭0 at 90 percent of maximum, which equals (0.9)(0.27) ⫽ 0.243. Hence, y1 ⫽ (2.2242, 0.6019, 0.1, 1.0994)T. Using the scaling matrix D, we can rescale the vector y1 back to the original variable as follows:

4.8

201


x1 ⫽ Dy1 ⫽ (8.8968, 0.6019, 0.5, 4.3976)T

and

Z1 ⫽ ⫺17.19

Thus, after one iteration, we have moved from x0 ⫽ (4, 1, 5, 4)T to the new interior point x1 ⫽ (8.8968, 0.6019, 0.5, 4.3976)T. The objective function value has been reduced from ⫺7.0 to ⫺17.19. Note that the optimal solution is (10, 0, 0, 5) and the minimum value of Z is ⫺20. The interior point algorithm will then rescale the point x1 by defining a new scaling matrix D so that x1 will be at unit distance from the coordinate axes of the new region. The process is continued until optimality is reached. Optimality Conditions

We use the weak duality theorem discussed in Section 4.8.1 to develop the optimality conditions for interior point algorithms. Consider the primal–dual problems given below: • Primal:

Minimize

Z ⫽ cTx

Subject to Ax ⫽ b

x⭓0

• Dual:

Maximize

bTw

Subject to ATw ⭐ c, w unrestricted in sign where w is an m-dimensional column vector of dual variables. The weak duality theorem states that if x0 is feasible for the primal and w0 is feasible for the dual, then cTx0 ⭓ bTw0. Moreover, when the objective values are equal, they are in fact optimal solutions to their respective problems. The difference between cTx0 and bTw0 is called the duality gap. When the duality gap is zero, we have the optimal solution. Thus, if we can get a feasible solution to the dual at each iteration, we can terminate the algorithm when the duality gap is nearly zero, that is, 兩cTx ⫺ bTw兩 ⬍ ␧

(4.21)

where ␧ is a very small number. As a matter of fact, we can get an estimate of the dual solution at each iteration using the direction vector d0. Using Eqs. (4.17) and (4.19), we can express d0 as follows:

202

LINEAR PROGRAMMING

d0 ⫽ ⫺Pp ⫽ ⫺[In ⫺ BT(BBT)⫺1B]p Using the relations B ⫽ AD and p ⫽ Dc, the above equation is reduced to d0 ⫽ ⫺D[c ⫺ AT(AD2AT)⫺1AD2c] Setting w0 ⫽ (AD2AT)⫺1AD2c

(4.22)

d0 ⫽ ⫺D(c ⫺ ATw0)

(4.23)

we get

From Eq. (4.23), if c ⫺ ATw0 ⭓ 0, then w0 is in fact a feasible solution to the dual. Otherwise, it is called an estimate of the dual solution. Using the estimate, we can check the optimality condition given by Eq. (4.21). In case c ⫺ ATw ⭓ 0 and the duality gap is nearly zero, we terminate the algorithm. The only step we have not yet addressed in the interior point method is how to find the initial interior point solution. One method is to use an approach similar to the Big M simplex method. We shall illustrate this using Example 4.12. Let us pick an arbitrary x0, say x0 ⫽ (10, 6, 1, 1)T, such that all its components are positive, but it may not satisfy the explicit constraints, Ax ⫽ b. Compute the infeasibility with respect to each constraint as follows: For constraint 1, the infeasibility equals 10 ⫺ x1 ⫺ x2 ⫺ x3 ⫽ ⫺7. For constraint 2, the infeasibility equals 5 ⫺ x2 ⫺ x4 ⫽ ⫺2. Introduce an artificial variable x5 whose constraint coefficients are equal to the infeasibility in each constraint. In our example, it will be ⫺7 and ⫺2, respectively. The artificial constraint system is then given by x1 ⫹ x2 ⫹ x3 x2

⫺ 7x5 ⫽ 10 ⫹ x4 ⫺ 2x5 ⫽ 5

Note that x0 ⫽ (10, 6, 1, 1, 1, 1)T is a feasible interior point to the artificial system. However, we want x5 to be removed as soon as possible by the interior point method. We can achieve that by assigning a large cost M to the artificial variable x5 in the objective function. Thus, the beginning LP problem to be solved becomes

4.8

Minimize


203

Z ⫽ ⫺2x1 ⫹ x2 ⫹ Mx5

Subject to x1 ⫹ x2 ⫹ x3 x2

⫺ 7x5 ⫽ 10 ⫹ x4 ⫺ 2x5 ⫽ 5

x1, x2, x3, x4, x5 ⭓ 0 with the initial interior point as x0 ⫽ (10, 6, 1, 1, 1)T. Note that if the artificial variable x5 cannot be driven to zero, the original LP problem is infeasible. Final Step. The interior point method will terminate with an ‘‘interior’’ so-

lution only. If one needs an extreme-point solution, a final step is necessary to convert the interior solution to a basic feasible solution. In the standard LP problem, there are m ⫹ n constraints, including the nonnegativity constraints, and n variables. A basic feasible solution must have at least n binding constraints, that is, at least n ⫺ m variables must be zero. If in the final solution there are less than n binding constraints, follow the steps given below: Step i. Determine a direction d in the null space of the binding constraints. Step ii. If cTd ⬍ 0, then move along the direction d (if cTd ⬎ 0, then move against the direction d), until one or more constraints become binding, that is, one or more variables become zero to satisfy nonnegativity conditions. Step iii. Repeat steps i and ii until n constraints are binding at the solution. In summary, the affine scaling algorithm operates directly on the standard LP problem and is less complicated. It is easy to implement and performs well in most problems. Unlike Karmarkar’s projective transformation method, it does not require that the optimal objective function value in the LP problem be zero nor does it require a ‘‘special structure,’’ that is, sum of all the variables equal to 1. Hence, it is one of the most popular interior point methods. However, the scaling algorithms (primal or dual) are sensitive to the proximity of the starting solution to the boundary of the feasible region. A primal–dual interior point algorithm employs a logarithmic barrier function to keep the interior point solutions sufficiently far away from the boundary. Current research indicates that the projective transformation methods do not compete favorably, during practical implementation, with affine scaling and primal– dual interior point algorithms. For an excellent discussion of affine scaling algorithms and other interior point methods, the reader is referred to the 1993 texts by Arbel [20] and Fang and Puthenpura [21]. Arbel’s text also contains a software package that implements interior point LP algorithms. The software can be run on IBM-PC or compatibles. Computational Efficiency of Interior Point Methods

Marsten et al. [18] have done extensive testing of the interior point methods. They claim that there is no fully robust implementation of Karmarkar’s pro-

204

LINEAR PROGRAMMING

jective transformation algorithm that is sufficiently efficient for testing against state-of-the-art simplex codes. Hence, most computer experiments have been done using codes implementing affine scaling algorithms and their variants, such as OSL and OB1. Marsten et al. [18] report that the number of iterations in the interior point method is almost independent of problem size. While solving a family of airline crew scheduling problems with 837 rows and columns varying between 50,000 and 4 million, they have observed that the number of iterations varied only between 20 and 100! In general, the number of iterations increases with the log of the number of variables, log n. However, it is important to recognize that as problem size grows, each iteration of the interior point method becomes very time consuming. In comparison, each iteration of the simplex algorithm (pivot step) is less cumbersome even though the number of iterations can be very large, usually less than 10m pivots. Marsten et al. [18] also compared the interior point method with the simplex method by solving a set of multiperiod oil refinery planning problems. This is a class of problems, as mentioned earlier in the section, that becomes particularly hard for the simplex method as the number of planning periods increases. They used OB1 (an interior point method) and XMP (the simplex method) for the study, both written by the same programmer in FORTRAN and run on the same computer, a DEC station 3100. For problems of size less than 2000 (i.e., m ⫹ n ⬍ 2000), the simplex method (XMP) performed better. For problems of size between 2000 and 5000, both simplex and interior point methods compete more evenly, with no single method showing a clear superiority. For problem sizes between 5000 and 20,000, OB1 was 2–10 times faster than XMP. However, Bixby [19] cautions users making any general conclusions on the superiority of the interior point methods at this time. His experiments seem to indicate that there is no single dominant method even for very large LP problems. For a detailed 1994 computational state-of-theart survey of interior point methods for linear programs, refer to Lustig et al. [22]. The main weakness of the interior point methods has to do with the ‘‘density’’ of the columns of the A matrix. Even a single column in A with all nonzero coefficients can make AAT completely dense, leading to enormous computational effort at each iteration. This does not happen in the revised simplex method since the sparsity of the original matrix is preserved throughout the solution process. It has been reported that the U.S. Military Airlift Command has been able to solve LP problems successfully with 150,000 variables and 12,000 constraints using interior point methods. Both American Air Lines and Delta Air Lines have been testing the use of interior point methods to solve very large LP problems for scheduling air crews and air crafts. Karmarkar’s algorithm and its many variants are current topics of research in the LP community. According to Fang and Puthenpura [21] and Lustig et al. [22], both approaches—simplex and interior point methods—are very sensitive to the problem structure and the performance depends on the sophisti-

4.8


205

cation of implementation. Hybrid methods using an interior point method at the beginning and then shifting to the simplex method, with an efficient basis recovery, are currently under investigation by Bixby et al. [23]. 4.8.4

Integer Programming

In practice, many LP problems do require integer solutions for some of the variables. For instance, it is not possible to employ a fractional number of workers or produce a fractional number of cars. The term (linear) integer programming refers to the class of LP problems wherein some or all of the decision variables are restricted to be integers. But solutions of integer programming problems are generally difficult, too time consuming, and expensive. Hence, a practical approach is to treat all integer variables as continuous and solve the associated linear program by the simplex method. We may be fortunate to get some of the values of the variables as integers automatically, but when the simplex method produces fractional solutions for some integer variables, they are generally rounded off to the nearest integer such that the constraints are not violated. This is very often used in practice, and it generally produces a good integer solution close to the optimal integer solution, especially when the values of the integer variables are large. There are situations when a model formulation may require the use of binary integer variables, which can take only the values 0 or 1. For these integer variables, rounding off produces poor integer solutions, and one does need other techniques to determine the optimal integer solution directly. Some of these techniques and many practical applications of integer programming are discussed in Chapter 11. 4.8.5

Goal Programming

All our discussion in this chapter considered only one objective function to optimize. However, in practical problems, one may desire to optimize several conflicting criteria simultaneously. One way of treating multiple criteria is to select one as the primary criterion and the rest as secondary criteria. The primary criterion is then used as the objective function to optimize, while the secondary criteria are assigned acceptable minimum or maximum values and are treated as problem constraints. However, if careful consideration were not given while selecting the acceptable levels, a feasible solution that will satisfy all the constraints may not exist. This problem is overcome by an approach called goal programming, which is fast becoming a practical method for handling conflicting multiple criteria. In goal programming, all the criteria are assigned acceptable target levels of achievement and a relative priority on achieving these levels. Goal programming treats these targets as goals to aspire for and not as absolute constraints. It then attempts to find a solution that comes as ‘‘close as possible’’ to the targets in the order of specified priorities. Details on how to formulate

206

LINEAR PROGRAMMING

goal programming models, their solution methods, and their applications are given in Chapter 11.

4.9

SUMMARY

In this chapter, we discussed linear programming, a special class of optimization problems with linear constraints and linear objective functions. Linear programming (LP) models are easy to solve computationally and have a wide range of applications in diverse fields. We also developed the simplex algorithm for solving general LP problems. We then discussed how to perform post-optimality analysis on the LP solutions. The ability to perform an analysis of this type is the main reason for the success of linear programming in practice. Other practical uses of linear programming are in solving integer and nonlinear programming problems as a sequence of LP problems.

REFERENCES 1. Gass, S. I., ‘‘Comments on the Possibility of Cycling with the Simplex Method,’’ Oper. Res., 27(4), 848–852 (1979). 2. Ravindran, A., D. T. Phillips and J. J. Solberg, Operations Research: Principles and Practice, 2nd ed., Wiley, New York, 1987. 3. Fourer, R., ‘‘Linear Programming Software Survey,’’ OR / MS Today, 30, 34–43 (Dec. 2003). 4. Brooke, A., D. Kendrick, and A Meeraus, GAMS: A User’s Guide, Scientific Press, Redwood City, CA, 1988. 5. Fourer, R., ‘‘Modelling Languages versus Matrix Generators for Linear Programming,’’ ACM Trans. Math Software, 9, 143–183 (1983). 6. Fourer, R., D. M. Gay, and B. W. Kernighan, ‘‘A Modeling Language for Mathematical Programming,’’ Manage. Sci., 36(5), 519–554 (1990). 7. Czyzyk J., J. H. Owen, and S. J. Wright, ‘‘Optimization on the Internet,’’ OR / MS Today, 24, 48–52 (Oct. 1997). 8. More, J. J., and S. J. Wright, Optimization Software Guide, SIAM Publications, Philadelphia, PA, 1993. 9. Khachian, L. G., ‘‘A Polynomial Algorithm in Linear Programming,’’ Soviet Math. Dokl., 20(1), 191–194 (1979). 10. Karmarkar, N. K., ‘‘A New Polynomial Time Algorithm for Linear Programming,’’ Combinatorica, 4, 373–395 (1984). 11. Felt, A., ‘‘LINDO: API: Software Review,’’ OR / MS Today, 29, 58–60 (Dec. 2002). 12. Gass, S. I., An Illustrated Guide to Linear Programming, McGraw-Hill, New York, 1970. 13. Salkin, H. M., and J. Saha, Studies in Linear Programming, North Holland / American Elsevier, New York, 1975.

PROBLEMS

207

14. Gass, S., Linear Programming, 4th ed., McGraw-Hill, New York, 1975. 15. Gray, P., and C. Cullinan-James, ‘‘Applied Optimization—A Survey,’’ Interfaces, 6(3), 24–41 (May 1976). 16. Dantzig, G. B., Linear Programming and Extensions, Princeton University Press, Princeton, NJ, 1963. 17. Murty, K. G., Operations Research: Deterministic Optimization Models, PrenticeHall, Englewood Cliffs, NJ, 1995. 18. Marsten, R. R., Subramanian, M. Saltzman, I. Lustig, and D. Shanno, ‘‘Interior Point Methods for Linear Programming: Just Call Newton, Lagrange, and Fiacco and McCormick,’’ Interfaces, 20(4), 105–116 (1990). 19. Bixby, R. E., ‘‘Progress in Linear Programming,’’ ORSA J. Comput., 6(1), 15–27 (1994). 20. Arbel, A., Exploring Interior Point Linear Programming: Algorithms and Software, MIT Press, Cambridge, MA, 1993. 21. Fang, S.-C., and S. Puthenpura, Linear Optimization and Extensions, PrenticeHall, Englewood Cliffs, NJ, 1993 22. Lustig, I. J., R. E. Marsten, and D. F. Shanno, ‘‘Interior Point Methods for Linear Programming: Computational State of the Art,’’ ORSA J. Comput., 6(1), 1–14 (1994). 23. Bixby, R. E., J. W. Gregory, I. J. Lusting, R. E. Marsten, and D. F. Shanno, ‘‘Very Large Scale Linear Programming: A Case Study in Combining Interior Point and Simplex Methods,’’ Operations Res., 40, 885–897 (1992).

PROBLEMS

4.1. What is the difference between the simplex method and exhaustive enumeration of feasible corner points of the constrained region? 4.2. What is the function of the minimum-ratio rule in the simplex method? 4.3. How do you recognize that an LP problem is unbounded while using the simplex method? 4.4. What is the function of the inner product rule in the simplex method? 4.5. What are artificial variables and why do we need them? How do they differ from slack–surplus variables? 4.6. Describe a systematic procedure for finding a feasible solution to a system of linear inequalities. 4.7. What is sensitivity analysis and why do we perform it? 4.8. What are shadow prices? What are their practical uses? 4.9. What are opportunity costs? 4.10. How does an addition of constant K to the objective function of an LP problem affect (i) its optimal solution and (ii) its optimal value?

208

LINEAR PROGRAMMING

4.11. Two products, A and B, are made that each involve two chemical operations. Each unit of product A requires 2 hr on operation 1 and 4 hr on operation 2. Each unit of product B requires 3 hr on operation 1 and 4 hr on operation 2. Available time for operation 1 is 16 hr, and for operation 2 it is 24 hr. The production of product B also results in a by-product C at no extra cost. Although some of this by-product can be sold at a profit, the remainder has to be destroyed. Product A sells for $4 profit per unit, while product B sells for $10 profit per unit. Byproduct C can be sold at a unit profit of $3, but if it cannot be sold, the destruction cost is $2 per unit. Forecasts show that up to five units of C can be sold. The company gets two units of C for each unit of B produced. Determine the production quantity of A and B, keeping C in mind, so as to make the largest profit. 4.12. Consider the problem of scheduling the weekly production of a certain item for the next 4 weeks. The production cost of the item is $10 for the first 2 weeks and $15 for the last 2 weeks. The weekly demands are 300, 700, 900, and 800 units, which must be met. The plant can produce a maximum of 700 units each week. In addition, the company can employ overtime during the second and third weeks. This increases the weekly production by an additional 200 units, but the cost of production increases by $5 per item. Excess production can be stored at a cost of $3 an item. How should the production be scheduled so as to minimize the total costs? Formulate this as an LP problem. 4.13. A refinery has four different crudes, which are to be processed to yield four products: gasoline, heating oil, jet fuel, and lube oil. There are maximum limits both on product demand (what can be sold) and crude availability. A schematic of the processing operation is given in Figure 4.10. Given the tabulated profits, costs, and yields (see Table 4.2):

Figure 4.10. Processing operation.

209

PROBLEMS

Table 4.2

Profits, Costs, Yields

4

Yields, bbl product / bbl crude Gasoline Heating oil Jet fuel Lube oil Othera Crude cost, $ / bbl Operating cost, $ / bbl a

1

2

3

Fuel Process

0.6 0.2 0.1 0 0.1 15.00 5.00

0.5 0.2 0.2 0 0.1 15.00 8.50

0.3 0.3 0.3 0 0.1 15.00 7.50

0.4 0.3 0.2 0 0.1 25.00 3.00

Lube Process

0.4 0.1 0.2 0.2 0.1 25.00 2.50

Product Value, $ / bbl

45.00 30.00 15.00 60.00

‘‘Other’’ refers to losses in processing.

(a) Set up the model appropriate for scheduling the refinery for maximum profit. (b) Suppose the given maximum product demands are changed to be minimum requirements. Set up the model appropriate for scheduling the refinery for minimum-cost operation. 4.14. A company manufactures three products: A, B, and C. Each unit of product A requires 1 hr of engineering service, 10 hr of direct labor, and 3 lb of material. To produce one unit of product B, it requires 2 hr of engineering, 4 hr of direct labor, and 2 lb of material. Each unit of product C requires 1 hr of engineering, 5 hr of direct labor, and 1 lb of material. There are 100 hr of engineering, 700 hr of labor, and 400 lb of material available. Since the company offers discounts for bulk purchases, the profit figures are as shown below: Product A

Product B

Product C

Sales, Units

Unit Profit, $

Sales, Units

Unit Profit, $

Sales Units,

Unit Profit, $

0–40 40–100 100–150 Over 150

10 9 8 7

0–50 50–100 Over 100

6 4 3

0–100 Over 100

5 4

For example, if 120 units of A are sold, the first 40 units will earn a profit of $10/unit, the next 60 will earn $9/unit, and the remaining 20

210

LINEAR PROGRAMMING

units $8/unit. Formulate a linear program to determine the most profitable product mix. 4.15. A scientist has observed a certain quantity Q as a function of a variable t. He is interested in determining a mathematical relationship relating t and Q, which takes the form Q(t) ⫽ at 3 ⫹ bt 2 ⫹ ct ⫹ d from the results of his n experiments (t1, Q1), (t2, Q2), . . . , (tn, Qn). He discovers that the values of the unknown coefficients a, b, c, and d must be nonnegative and should add up to 1. To account for errors in his experiments, he defines an error term, ei ⫽ Qi ⫺ Q(ti) He wants to determine the best values for the coefficients a, b, c, and d using the following criterion functions:

冘兩e 兩 n

Criterion 1:

Minimize Z ⫽

i

i⫽1

Criterion 2:

Minimize

max 兩ei兩 i

where ei is the value of the error associated with the ith experiment. Show that the scientist’s problem reduces to an LP problem under both criterion 1 and criterion 2. 4.16. A machine shop has one drill press and five milling machines, which are to be used to produce an assembly consisting of parts 1 and 2. The productivity of each machine for the two parts is given below: Production Time, min / piece Part

Drill

Mill

1 2

3 5

20 15

It is desired to maintain a balanced loading on all machines such that no machine runs more than 30 min/day longer than any other machine (assume that the milling load is split evenly among all five milling machines). Divide the work time of each machine to obtain the maximum number of completed assemblies, assuming an 8-hr working day.

PROBLEMS

211

4.17. A company makes a specialty solvent at two levels of purity, which it sells in gallon containers. Product A is of higher purity than product B, and profits are $0.40/gal on A and $0.30/gal on B. Product A requires twice the processing time of B. If the company produced only B, it could make 100 gal/day. However, process throughput limitations permit production of only 800 gal/day of both A and B combined. Contract sales require that at least 200 gal/day of B be produced. Assuming all of the product can be sold, what volumes of A and B should be produced? Solve using the tableau form of the simplex method. Confirm your solution using graphical means. 4.18. Transform the following linear program to the standard form: Minimize Subject to

Z ⫽ ⫺3x1 ⫹ 4x2 ⫺ 2x3 ⫹ 5x4 4x1 ⫺ x2 ⫹ 2x3 ⫺ x4 ⫽ ⫺2 x1 ⫹ x2 ⫹ 3x3 ⫺ x4 ⭐ 14 ⫺2x1 ⫹ 3x2 ⫺ x3 ⫹ 2x4 ⭓ 2

x1, x2 ⭓ 0

x3 ⭐ 0

x4 unrestricted in sign

4.19. Consider a system of two equations in five unknowns as follows: x1 ⫹ 2x2 ⫹ 10x3 ⫹ 4x4 ⫺ 2x5 ⫽ 5 x1 ⫹ x2 ⫹ 4x3 ⫹ 3x4 ⫹ x5 ⫽ 8 x1, . . . , x5 ⭓ 0 (a) Reduce the system to canonical form with respect to (x1, x2) as basic variables. Write down the basic solution. Is it feasible? Why or why not? (b) What is the maximum number of basic solutions possible? (c) Find a canonical system that will give a basic feasible solution to the above system by trial and error. 4.20. Consider the following linear program: Maximize

Z ⫽ 2x1 ⫺ x2 ⫹ x3 ⫹ x4

Subject to ⫺x1 ⫹ x2 ⫹ x3 x1 ⫹ x2

⫹x5 ⫹x4

2x1 ⫹ x2 ⫹ x3 x1, . . . , x6 ⭓ 0

⫽1 ⫽2 ⫹ x6 ⫽ 6

212

LINEAR PROGRAMMING

(a) Write down the initial basic feasible solution by inspection. (b) Find a feasible solution by increasing the nonbasic variable x1 by one unit while holding x2 and x3 as zero. What will be the net change in the objective function? (c) What is the maximum possible increase in x1 subject to the constraints? (d) Find the new basic feasible solution when x1 is increased to its maximum value found in (c). (e) Is the new basic feasible solution obtained in (d) optimal? Why or why not? 4.21. Use the simplex method to solve Maximize

Z ⫽ x1 ⫹ 3x2

Subject to x1

⭐5

x1 ⫹ 2x2 ⭐ 10 x2 ⭐ 4 x1, x2 ⭓ 0 Plot the feasible region using x1 and x2 as coordinates. Follow the solution steps of the simplex method graphically by interpreting the shift from one basic feasible solution to the next in the feasible region. 4.22. Find an optimal solution to the following linear program by inspection:

Minimize

Z ⫽ x1 ⫺ 3x2 ⫹ 2x3

Subject to ⫺2 ⭐ x1 ⭐ 3 0 ⭐ x2 ⭐ 4

2 ⭐ x3 ⭐ 5

4.23. Use the simplex method to solve Minimize

Z ⫽ 3x1 ⫹ x2 ⫹ x3 ⫹ x4

Subject to ⫺2x1 ⫹ 2x2 ⫹ x3 ⫽ 4 3x1 ⫹ x2 ⫹ x4 ⫽ 6 x1, x2, x3, x4 ⭓ 0 Find an alternative optimal solution if one exits.

213

PROBLEMS

4.24. Use the simplex method to solve Maximize Subject to

Z ⫽ x1 ⫹ 2x2 ⫹ 3x3 ⫹ 4x4 x1 ⫹ 2x2 ⫹ 2x3 ⫹ 3x4 ⭐ 20 2x1 ⫹ x2 ⫹ 3x3 ⫹ 2x4 ⭐ 20 x1, . . . , x4 ⭓ 0

Is the optimal solution unique? Why or why not? 4.25. Use the simplex method to verify that the following problem has no optimal solution: Maximize

Z ⫽ x1 ⫹ 2x2

Subject to ⫺2x1 ⫹ x2 ⫹ x3 ⭐ 2 ⫺x1 ⫹ x2 ⫺ x3 ⭐ 1

x1, x2, x3 ⭓ 0 From the final simplex tableau, construct a feasible solution whose value of the objective function is greater than 2000. 4.26. Consider the standard LP problem Z ⫽ cx

Minimize

Subject to Ax ⫽ b

x⭓0

Let the vectors x (1) and x (2) be two optimal solutions to the above problem. Show that the vector x(␭) ⫽ ␭x (1) ⫹ (1 ⫺ ␭)x (2) is also an optimal solution for any value of ␭ between 0 and 1. (Note: The above result is very useful in linear programming. Once we have two optimal solutions to a linear program, then we can generate an infinite number of optimal solutions by varying ␭ between 0 and 1.) 4.27. In the tableau for the maximization problem below, the values of the six constants ␣1, ␣2, ␣3, ␤, ␳1, ␳2 are unknown (assume there are no artificial variables): Basis

x1

x2

x3

x4

x5

x6

Constants

x3 x4 x6

4 ⫺1 ␣3

␣1 ⫺5 ⫺3

1 0 0

0 1 0

␣2 ⫺1 ⫺4

0 0 1

␤ 2 3

c row

␳1

␳2

0

0

⫺3

0

x1, . . . , x6 ⭓ 0

214

LINEAR PROGRAMMING

State restrictions on the six unknowns (␣1, ␣2, ␣3, ␤, ␳1, ␳2) that would make the following statements true about the given tableau: (a) The current solution is optimal but an alternate optimum exists. (b) The current solution is infeasible. (State which variable.) (c) One of the constraints is inconsistent. (d) The current solution is a degenerate basic feasible solution. (Which variable causes degeneracy?) (e) The current solution is feasible but the problem has no finite optimum. (f) The current solution is the unique optimum solution. (g) The current solution is feasible but the objective can be improved by replacing x6 by x1. What will be the total change in the objective function value after the pivot? 4.28. A company that manufacturers three products A, B, and C using three machines M1, M2, and M3 wants to determine the optimal production schedule that maximizes the total profit. Product A has to be processed by machines M1, M2, and M3; product B requires M1 and M3; while product C requires M1 and M2. The unit profits on the three products are $4, $2, and $5, respectively. The following linear program is formulated to determine the optimal product mix: Maximize Subject to

Z ⫽ 4x1 ⫹ 2x2 ⫹ 5x3 x1 ⫹ 2x2 ⫹ x3 ⭐ 430

(machine 1)

⫹ 2x3 ⭐ 460

(machine 2)

⭐ 450

(machine 3)

3x1 x1 ⫹ 4x2

x1, x2, x3 ⭓ 0 where x1, x2, and x3 are the amounts of products A, B, and C and the constraints reflect the available capacities of M1, M2, and M3. The computer prints out the following solution: x1 ⫽ 0, x2 ⫽ 100, x3 ⫽ 230 max Z ⫽ 1350 1.0, 2.0, and 0.0 for constraints 1, 2, and 3, respectively Opportunity costs 3.0, 0, and 0 for variables x1, x2, x3, respectively Optimal solution Optimal value Shadow prices

215

PROBLEMS

Ranges on Objective Function Coefficients Variable

Lower Limit

Present Value

Upper Limit

x1 x2 x3

⫺⬁

0 2.333

4.0 2.0 5.0

7.0 10.0 ⬁

Ranges on RHS Constants Row

Lower Limit

Present Value

Upper Limit

1 2 3

230 440 400

430 460 450

440 860 ⬁

(a) Because of an increase in the cost of raw material used for product C, its unit profit drops to $4. Determine the new optimal solution and the maximum profit. (b) Suppose it is possible to increase the capacity of one of the machines. Which one would you recommend for expansion and why? (c) Due to an improvement in product design, the unit profit on product A can be increased to $6. Is it worthwhile to produce product A now? Explain. (d) Suppose the capacity of machine 2 can be increased by another 200 min at a cost of $250. Is it economical to do so? Explain. 4.29. Write down the dual of the following linear programs: (a) Minimize Z ⫽ x1 ⫹ 2x2 ⫹ 3x3 ⫹ 4x4 Subject to

x1 ⫹ 2x2 ⫹ 2x3 ⫹ 3x4 ⭌ 30 2x1 ⫹ x2 ⫹ 3x3 ⫹ 2x4 ⭌ 20 x1, x2, x3, x4 ⭌ 0

(b)

Maximize

Z ⫽ ⫺4x1 ⫺ 3x2

Subject to

x1 ⫹ x2 ⬉ 1 ⫺ x2 ⬉ ⫺1 ⫺x1 ⫹ 2x2 ⬉ 1

x1, x2 ⭌ 0

216

LINEAR PROGRAMMING

(c)

Minimize Subject to

Z ⫽ 40x1 ⫹ 36x2 x1

⬉8

x2 ⬉ 10 5x1 ⫹ 3x2 ⭌ 45 x1, x2 ⭌ 0

4.30. Consider the following linear program which we call the primal problem: Maximize

Z ⫽ 3x1 ⫹ 2x2

Subject to ⫺x1 ⫹ 2x2 ⬉ 4 3x1 ⫹ 2x2 ⬉ 14 x1 ⫺ x2 ⬉ 3 x1, x2 ⭌ 0 (a) Write down the dual problem. (b) Prove using the duality theory that there exists an optimal solution to both the primal and dual problems. (c) Compute an upper and a lower bound on the optimal values of both the problems using the weak duality theorem. 4.31. Prove using the duality theory that the following linear program is feasible but has no optimal solution: Minimize

Z ⫽ x1 ⫺ x2 ⫹ x3

Subject to x1

⫺ x3 ⭌ 4

x1 ⫺ x2 ⫹ 2x3 ⭌ 3 x1, x2 x3 ⭌ 0 4.32. Consider Example 4.12 discussed in Section 4.8.3. Determine a direction of decrease that will be feasible at the following points: (a) x1 ⫽ (5, 5, 0, 0) (b) x2 ⫽ (1, 5, 4, 0) (c) x3 ⫽ (0, 0, 10, 5)

PROBLEMS

217

4.33. Consider the LP problem given below: Maximize

Z ⫽ x1 ⫹ x2

Subject to

2x1 ⫹ x2 ⭐ 24 x2 ⭐ 8 x1, x2 ⭓ 0

Given the initial interior point as x1 ⫽ 2 and x2 ⫽ 4, determine the next interior point using the affine scaling algorithm. Show the path taken by the interior point method on a graph.

5 CONSTRAINED OPTIMALITY CRITERIA In Chapters 2 and 3, we discussed the necessary and sufficient optimality criteria for unconstrained optimization problems. But most engineering problems involve optimization subject to several constraints on the design variables. The presence of constraints essentially reduces the region in which we search for the optimum. At the outset, it may appear that the reduction in the size of the feasible region should simplify the search for the optimum. On the contrary, the optimization process becomes more complicated, since many of the optimality criteria developed earlier need not hold in the presence of constraints. Even the basic condition that an optimum must be at a stationary point, where the gradient is zero, may be violated. For example, the unconstrained minimum of ƒ(x) ⫽ (x ⫺ 2)2 occurs at the stationary point x ⫽ 2. But if the problem has a constraint that x ⭓ 4, then the constrained minimum occurs at the point x ⫽ 4. This is not a stationary point of ƒ, since ƒ⬘(4) ⫽ 4. In this chapter we develop necessary and sufficient conditions of optimality for constrained problems. We begin with optimization problems involving only equality constraints.

5.1

EQUALITY-CONSTRAINED PROBLEMS

Consider an optimization problem involving several equality constraints: Minimize

ƒ(x1, x2, . . . , xN)

Subject to hk (x1, . . . , xN) ⫽ 0

218

k ⫽ 1, . . . , K


5.2

LAGRANGE MULTIPLIERS

219

In principle, this problem can be solved as an unconstrained optimization problem by explicitly eliminating K independent variables using the equality constraints. In effect, the presence of equality constraints reduces the dimensionality of the original problem from N to N ⫺ K. Once the problem is reduced to an unconstrained optimization problem, the methods of Chapters 2 and 3 can be used to identify the optimum. To illustrate this, consider the following example. Example 5.1

Minimize

ƒ(x) ⫽ x1x2x3

Subject to h1(x) ⫽ x1 ⫹ x2 ⫹ x3 ⫺ 1 ⫽ 0 Eliminating the variable x3 with the help of h1(x) ⫽ 0, we get an unconstrained optimization problem involving two variables: min ƒ(x1, x2) ⫽ x1x2(1 ⫺ x1 ⫺ x2) The methods of Chapter 3 can now be applied to identify the optimum. The variable elimination method is applicable as long as the equality constraints can be solved explicitly for a given set of independent variables. In the presence of several equality constraints, the elimination process may become unwieldy. Moreover, in certain situations it may not be possible to solve the constraints explicitly to eliminate a variable. For instance, in Example 5.1, if the constraints h1(x) ⫽ 0 were given as h1(x) ⫽ x21x3 ⫹ x2x23 ⫹ x⫺1 2 x1 ⫽ 0 then no explicit solution of one variable in terms of the others would be possible. Hence, in problems involving several complex equality constraints, it is better to use the method of Lagrange multipliers, which is described in the next section, for handling the constraints.

5.2


The method of Lagrange multipliers essentially gives a set of necessary conditions to identify candidate optimal points of equality-constrained optimization problems. This is done by converting the constrained problem to an equivalent unconstrained problem with the help of certain unspecified parameters known as Lagrange multipliers.

220

CONSTRAINED OPTIMALITY CRITERIA

Consider the minimization of a function of n variables subject to one equality constraint: Minimize

ƒ(x1, x2, . . . , xN)

Subject to h1(x1, x2, . . . , xN) ⫽ 0

(5.1) (5.2)

The method of Lagrange multipliers converts this problem to the following unconstrained optimization problem: Minimize L(x, v) ⫽ ƒ(x) ⫺ vh1(x)

(5.3)

The unconstrained function L(x; v) is called the Lagrangian function, and v is an unspecified constant called the Lagrange multiplier. There are no sign restrictions on the value of v. Suppose for a given fixed value of v ⫽ vo the unconstrained minimum of L(x; v) with respect to x occurs at x ⫽ x o and x o satisfies h1(x o) ⫽ 0. then, it is clear that x o minimizes Eq. (5.1) subject to (5.2) because for all values of x that satisfy (5.2), h1(x) ⫽ 0 and min L(x; v) ⫽ min ƒ(x). Of course, the challenge is to determine the appropriate value for v ⫽ vo so that the unconstrained minimum point x o satisfies (5.2). But this can be done by treating v as a variable, finding the unconstrained minimum of (5.3) as a function of v, and adjusting v such that (5.2) is satisfied. We illustrate this with the following example: Example 5.2

Minimize

ƒ(x) ⫽ x21 ⫹ x22

Subject to h1(x) ⫽ 2x1 ⫹ x2 ⫺ 2 ⫽ 0 The unconstrained minimization problem becomes Minimize L(x; v) ⫽ x21 ⫹ x22 ⫺ v(2x1 ⫹ x2 ⫺ 2) Solution. Setting the gradient of L with respect to x equal to zero, ⭸L ⫽ 2x1 ⫺ 2v ⫽ 0 ⇒ x 1o ⫽ v ⭸x1 ⭸L v ⫽ 2x2 ⫺ v ⫽ 0 ⇒ x o2 ⫽ ⭸x2 2

To test whether the stationary point x o corresponds to a minimum, we compute the Hessian matrix of L(x; v) with respect to x as

5.2

HL(x; v) ⫽


221

冋册 2 0

0 2

which is positive definite. This implies that HL(x; v) is a convex function for all x. Hence x o1 ⫽ v, x o2 ⫽ v /2 corresponds to the global minimum. To determine the optimal v, we substitute the values of x o1 and x o2 in the constraint 2x1 ⫹ x2 ⫽ 2 to get 2v ⫹ v /2 ⫽ 2 or vo ⫽ –45 . Hence the constrained minimum is attained at x o1 ⫽ –45 , x o2 ⫽ –25 , and min ƒ(x) ⫽ –45 . In the solution of Example 5.2, we treated L(x; v) as a function of two variables x1 and x2 and considered v as a parameter whose value is ‘‘adjusted’’ to satisfy the constraint. In problems where it is difficult to get an explicit solution to ⭸L ⫽0 ⭸xj

for j ⫽ 1, 2, . . . , N

as a function of v, the values of x and v can be determined simultaneously by solving the following system of N ⫹ 1 equations in N ⫹ 1 unknowns: ⭸L ⫽0 ⭸xj

for j ⫽ 1, 2, . . . , N

h1(x) ⫽ 0 Any appropriate numerical search technique discussed in Chapter 3 (e.g., Newton’s method) could be used to determine all possible solutions. For each of the solutions (x o; vo), the Hessian matrix of L with respect to x has to be evaluated to determine whether it is positive definite for a local minimum (or negative definite for a local maximum). Example 5.3

Maximize

ƒ(x) ⫽ x1 ⫹ x2

Subject to x21 ⫹ x22 ⫽ 1

222


Solution L(x; v) ⫽ x1 ⫹ x2 ⫺ v(x12 ⫹ x22 ⫺ 1) ⭸L ⫽ 1 ⫺ 2vx1 ⫽ 0 ⭸x1 ⭸L ⫽ 1 ⫺ 2vx2 ⫽ 0 ⭸x2

h1(x) ⫽ x21 ⫹ x22 ⫺ 1 ⫽ 0 There are two solutions to this system of three equations in three variables, given by

冉冉

冪冊

1 1 1 ,⫺ ;⫺ 2 兹2 兹2

(x (1); v1) ⫽ ⫺ (x (2); v2) ⫽

1 1 , ; 兹2 兹2

冪21冊

The Hessian matrix of L(x; v) with respect to x is given by HL(x; v) ⫽

冋

⫺2v

0

0

册

⫺2v

Evaluating the matrix H at the two solutions, we find that

冋

兹2

HL(x (1); v1) ⫽

0

0 兹2

册

is positive definite

and HL(x (2); v2) ⫽

冋

⫺兹2

0

册

0 ⫺兹2

is negative definite

Hence, (x (2); v2) corresponds to the maximum of L with respect to x, and the optimal solution is x o1 ⫽ x o2 ⫽ 1/ 兹2. [Note that (x (1); v1) corresponds to the minimum of L.] It is to be emphasized here that if we consider L as a function of three variables, namely, x1, x2 and v, then the points (x (1); v1) and (x (2); v2) do not correspond to a minimum or maximum of L with respect to x and v. As a matter of fact, they become saddlepoints of the function L(x, v). We shall discuss saddlepoints and their significance later in this chapter.

5.2


223

The Lagrange multiplier method can be extended to several equality constraints. Consider the general problem Minimize

ƒ(x)

Subject to hk (x) ⫽ 0

k ⫽ 1, 2, . . . , K

The Langrange function becomes

冘vh K

L(x; v) ⫽ ƒ(x) ⫺

k

k

k⫽1

Here v1, v2, . . . , vK are the Lagrange multipliers, which are unspecified parameters whose values will be determined later. By setting the partial derivatives of L with respect to x equal to zero, we get the following system of N equations in N unknowns: ⭸L(x; v) ⫽0 ⭸x1 ⭸L(x; v) ⫽0 ⭸x2 ⯗ ⭸L(x; v) ⫽0 ⭸xN

It is difficult to solve the above system as a function of the vector v explicitly, we can augment the system with the constraint equations h1(x) ⫽ 0 h2(x) ⫽ 0 ⯗

hK (x) ⫽ 0 A solution to the augmented system of N ⫹ K equations in N ⫹ K variables gives a stationary point of L. This can then be tested for minimum or maximum by computing the Hessian matrix of L with respect to x as discussed in the single-constraint case. There may exist some problems for which the augmented system of N ⫹ K equations in N ⫹ K unknowns may not have a solution. In such cases, the Lagrange multiplier method would fail. However, such cases are rare in practice.

224

5.3


ECONOMIC INTERPRETATION OF LAGRANGE MULTIPLIERS

So far in our discussion we have treated Lagrange multipliers as adjustable parameters whose values are adjusted to satisfy the constraints. In fact, the Lagrange multipliers do have an important economic interpretation as shadow prices of the constraints, and their optimal values are very useful in sensitivity analysis. To exhibit this interpretation, let us consider the following optimization problem involving two variables and one equality constraint: Minimize

ƒ(x1, x2)

Subject to h1(x1, x2) ⫽ b1 where b1 corresponds to the availability of a certain scarce resource, L(x; v1) ⫽ ƒ(x) ⫺ v1[h1(x) ⫺ b1] Let us assume that the stationary point of L corresponds to the global minimum: ⭸L ⭸ƒ ⭸h ⫽ ⫺ v1 1 ⫽ 0 ⭸x1 ⭸x1 ⭸x1

(5.4)

⭸L ⭸ƒ ⭸h ⫽ ⫺ v1 1 ⫽ 0 ⭸x2 ⭸x2 ⭸x2

(5.5)

Let vo1 be the optimal Lagrange multiplier and x o be the optimal solution. Let the minimum of L(x; v) for v1 ⫽ v1o occur at x ⫽ x o such that h1(x o) ⫽ b1 and ƒ(x o) ⫽ L(x o; vo1) ⫽ ƒo. It is clear that the optimal values (x o; vo1) are a function of b1, the limited availability of the scarce resource. The change in ƒo, the optimal value of ƒ, due to a change in b1 is given by the partial derivative ⭸ƒo / ⭸b1. By the chain rule, ⭸ƒo ⭸ƒo ⭸x o ⭸ƒo ⭸x o ⫽ o䡠 1⫹ o䡠 2 ⭸b1 ⭸x 1 ⭸b1 ⭸x 2 ⭸b1

(5.6)

The partial derivative of the constraint function h1(x) ⫺ b1 ⫽ 0 is given by ⭸h1 ⭸x 1o ⭸h ⭸x o 䡠 ⫹ 01 䡠 2 ⫺ 1 ⫽ 0 o ⭸x 1 ⭸b1 ⭸x2 ⭸b1

Multiply both sides of (5.7) by vo1 and subtract from (5.6) to get

(5.7)

5.4

⭸ƒo ⫽ vo1 ⫹ ⭸b1

KUHN–TUCKER CONDITIONS

冘冉⭸⭸xƒ ⫺ v ⭸⭸xh 冊 ⭸⭸xb 2

o

j⫽1

o 1

o j

1 o j

o j

225

(5.8)

1

Since x o and vo1 satisfy (5.4) and (5.5), Eq. (5.8) reduces to ⭸ƒo ⫽ vo1 ⭸b1

(5.9)

Thus, from (5.9) we note that the rate of change of the optimal value of ƒ with respect to b1 is given by the optimal value of the Lagrange multiplier vo1. In other words, the change in the optimal value of the objective function per unit increase in the right-hand-side constant of a constraint, which we defined as shadow price in Chapter 4, is given by the Lagrange multiplier. Depending on the sign of vo1, ƒo may increase or decrease with a change in b1. For an optimization problem with K constraints and n variables given by Minimize

ƒ(x)

Subject to hk (x) ⫽ bk

k ⫽ 1, 2, . . . , K

we can show, using a similar argument, that ⭸ƒo ⫽ vok ⭸bk

5.4

for k ⫽ 1, 2, . . . , K


In the previous section we found that Lagrange multipliers could be used in developing optimality criteria for the equality-constrained optimization problems. Kuhn and Tucker have extended this theory to include the general nonlinear programming (NLP) problem with both equality and inequality constraints. Consider the following general NLP problem: Minimize

ƒ(x)

(5.10)

Subject to

gj (x) ⭓ 0

for j ⫽ 1, 2, . . . , J

hk (x) ⫽ 0

for k ⫽ 1, 2, . . . , K

x ⫽ (x1, x2, . . . , xN)

(5.11) (5.12)

226


Definition The inequality constraint gj (x) ⭓ 0 is said to be an active or binding constraint at the point x if gj (x) ⫽ 0; it is said to be inactive or nonbinding if gj (x) ⬎ 0. If we can identify the inactive constraints at the optimum before solving the problem, then we can delete those constraints from the model and reduce the problem size. The main difficulty lies in identifying the inactive constraints before the problem is solved. Kuhn and Tucker have developed the necessary and sufficient optimality conditions for the NLP problem assuming that the functions ƒ, gj, and hk are differentiable. These optimality conditions, commonly known as the Kuhn– Tucker conditions (KTCs) may be stated in the form of finding a solution to a system of nonlinear equations. Hence, they are also referred to as the Kuhn– Tucker problem (KTP).

5.4.1

Kuhn–Tucker Conditions or Kuhn–Tucker Problem

Find vectors x(N⫻1), u(1⫻J), and v(1⫻K) that satisfy

冘 u ⵜg (x) ⫺ 冘 v ⵜh (x) ⫽ 0 J

ⵜƒ(x) ⫺

K

j

j

k

j⫽1

k

(5.13)

k⫽1

gj (x) ⭓ 0

for j ⫽ 1, 2, . . . , J

(5.14)

hk (x) ⫽ 0

for k ⫽ 1, 2, . . . , K

(5.15)

ujgj (x) ⫽ 0

for j ⫽ 1, 2, . . . , J

(5.16)

uj ⭓ 0

for j ⫽ 1, 2, . . . , J

(5.17)

Let us first illustrate the KTCs with an example. Example 5.4

Minimize

ƒ(x) ⫽ x21 ⫺ x2

Subject to x1 ⫹ x2 ⫽ 6 x1 ⫺ 1 ⭓ 0

x21 ⫹ x22 ⭐ 26

5.4


227

Solution. Expressing the above problem in the NLP problem format given by Eqs. (5.12) and (5.13), we get ƒ(x) ⫽ x21 ⫺ x2

ⵜƒ(x) ⫽ (2x1, ⫺ 1)

g1(x) ⫽ x1 ⫺ 1

ⵜg1(x) ⫽ (1, 0)

g2(x) ⫽ 26 ⫺ x ⫺ x

ⵜg2(x) ⫽ (⫺2x1, ⫺2x2)

h1(x) ⫽ x1 ⫹ x2 ⫺ 6

ⵜh1(x) ⫽ (1, 1)

2 1

2 2

Equation (5.13) of the KTCs reduces to ⭸ƒ ⭸g ⭸g ⭸h ⫺ u1 1 ⫺ u2 2 ⫺ v1 1 ⫽ 0 ⭸xj ⭸xj ⭸xj ⭸xj

for j ⫽ 1, 2

This corresponds to 2x1 ⫺ u1 ⫹ 2x1u2 ⫺ v1 ⫽ 0 ⫺1 ⫹ 2x2u2 ⫺ v1 ⫽ 0

Equations (5.14) and (5.15) of the KTP correspond to the given constraints of the NLP problem and are given by x1 ⫺ 1 ⭓ 0 26 ⫺ x12 ⫺ x22 ⭓ 0 x1 ⫹ x2 ⫺ 6 ⫽ 0 Equation (5.16) is known as the complementary slackness condition in the KTP and is given by u1(x1 ⫺ 1) ⫽ 0 u2(26 ⫺ x12 ⫺ x22) ⫽ 0 Note that the variables u1 and u2 are restricted to be zero or positive, while v1 is unrestricted in sign. Thus, the KTCs for this example are given by

228


2x1 ⫺ u1 ⫹ 2x1u2 ⫺ v1 ⫽ 0 ⫺1 ⫹ 2x2u2 ⫺ v1 ⫽ 0

x1 ⫺ 1 ⭓ 0 26 ⫺ x21 ⫺ x22 ⭓ 0 x1 ⫹ x2 ⫺ 6 ⫽ 0 u1(x1 ⫺ 1) ⫽ 0 u2(26 ⫺ x12 ⫺ x22) ⫽ 0 u1 ⭓ 0 5.4.2

u2 ⭓ 0

v1 unrestricted

Interpretation of Kuhn–Tucker Conditions

To interpret the KTCs, consider the equality-constrained NLP problem Minimize

ƒ(x)


k ⫽ 1, . . . , K

The KTCs are given by ⵜƒ(x) ⫺

冘 v ⵜh (x) ⫽ 0

(5.18)

hk (x) ⫽ 0

(5.19)

k

k

k

Now consider the Lagrangian function corresponding to the equalityconstrained NLP problem: L(x; v) ⫽ ƒ(x) ⫺

冘 v h (x) k

k

k

The first-order optimality conditions are given by ⵜL(x) ⫽ ⵜƒ(x) ⫺

冘 v ⵜh (x) ⫽ 0 k

k

k

ⵜL(v) ⫽ hk (x) ⫽ 0

We find that the KTCs (5.18) and (5.19) are simply the first-order optimality conditions of the Lagrangian problem. Let us consider the inequality-constrained NLP problem:

5.5

Minimize

KUHN–TUCKER THEOREMS

229

ƒ(x)

Subject to gj (x) ⭓ 0

j ⫽ 1, . . . , J

The KTCs are given by ⵜƒ(x) ⫺

冘 u ⵜg (x) ⫽ 0 j

j

j

gj (x) ⭓ 0

ujgj (x) ⫽ 0

uj ⭓ 0

The Lagrangian function can be expressed as L(x; u) ⫽ ƒ(x) ⫺

冘 u g (x) j j

j

the first-order optimality conditions are given by ⵜƒ(x) ⫺

冘 u ⵜg (x) ⫽ 0 j

j

(5.20)

j

gj (x) ⭓ 0

for j ⫽ 1, . . . , J

Note that uj is the Lagrange multiplier corresponding to constraint j. In Section 5.3 we showed that uj represents the shadow price of constraint j; in other words, uj gives the change in the minimum value of the objective function ƒ(x) per unit increase in the right-hand-side constant. If we assume that the jth constraint is inactive [i.e., gj (x) ⬎ 0], then uj ⫽ 0 and ujgj (x) ⫽ 0. On the other hand, if the jth constraint is active [i.e., gj (x) ⫽ 0], then its shadow price uj need not necessarily be zero, but the value ujgj (x) ⫽ 0 since gj (x) ⫽ 0. Hence, ujgj (x) ⫽ 0

for all j ⫽ 1, . . . , J

To determine the sign of uj or the sign of the shadow price of the constraint gj (x) ⭓ 0, let us increase its right-hand-side value from 0 to 1. It is clear that this will constrain the problem further, because any solution that satisfied gj (x) ⭓ 1 will automatically satisfy gj (x) ⭓ 0. Hence, the feasible region will become smaller, and the minimum value of ƒ(x) cannot improve (i.e., will generally increase). In other words, the shadow price of the jth constraint uj is nononegative, as given by the KTCs.

5.5


In the previous section we developed the KTCs for constrained optimization problems. Using the theory of Lagrange multipliers, we saw intuitively that

230


the KTCs give the necessary conditions of optimality. In this section, we see the precise conditions under which the KTP implies the necessary and sufficient conditions of optimality. Theorem 5.1

Kuhn–Tucker Necessity Theorem

Consider the NLP problem given by Eqs. (5.10)–(5.12). Let ƒ, g, and h be differentiable functions and x* be a feasible solution to the NLP problem. Let I ⫽ {j兩gj (x*) ⫽ 0}. Futhermore, ⵜgj (x*) for j 僆 I and ⵜhk (x*) for k ⫽ 1, . . . , K are linearly independent. If x* is an optimal solution to the NLP problem, then there exists a (u*, v*) such that (x*, u*, v*) solves the KTP given by Eqs. (5.13)–(5.17). The proof of the theorem is beyond the scope of this test. Interested students may refer to Bazaraa et al. [1, Chap. 4]. The conditions that ⵜgj (x*) for j 僆 I and ⵜhk (x*) for k ⫽ 1, . . . , K are linearly independent at the optimum are known as a constraint qualification. A constraint qualification essentially implies certain regularity conditions on the feasible region that are frequently satisfied in practical problems. However, in general, it is difficult to verify the constraint qualifications, since it requires that the optimum solution be known beforehand. For certain special NLP problems, however, the constraint qualification is always satisfied: 1. When all the inequality and equality constraints are linear. 2. When all the inequality constraints are concave functions and the equality constraints are linear and there exists at least one feasible x that is strictly inside the feasible region of the inequality constraints. In other words, there exists an x such that gj(x) ⬎ 0 for j ⫽ 1, . . . , J and hk(x) ⫽ 0 for k ⫽ 1, . . . , K. When the constraint qualification is not met at the optimum, there may not exist a solution to the KTP. Example 5.5

Minimize

ƒ(x) ⫽ (x1 ⫺ 3)2 ⫹ x22

Subject to g1(x) ⫽ (1 ⫺ x1)3 ⫺ x2 ⭓ 0 g2(x) ⫽ x1 ⭓ 0 g3(x) ⫽ x2 ⭓ 0

5.5

231


Solution. Figure 5.1 illustrates the feasible region for this nonlinear program. It is clear that the optimal solution to this problem is x*1 ⫽ 1, x*2 ⫽ 0, and ƒ(x*) ⫽ 4. We shall now show that the constraint qualification is not satisfied at the optimum. Since g1(x*) ⫽ 0, g2(x*) ⬎ 0, and g3(x*) ⫽ 0, I ⫽ {1, 3}. Now, ⵜg1(x*) ⫽ [⫺3(1 ⫺ x1)2, ⫺ 1]x⫽x* ⫽ (0, ⫺1)

and ⵜg3(x*) ⫽ (0, 1)

It is clear that ⵜg1(x*) and ⵜg3(x*) are not linearly independent. Hence, the constraint qualification is not satisfied at the point x* ⫽ (1, 0). Let us now write the KTCs to see whether they will be satisfied at (1, 0). Equations (5.13), (5.16), and (5.17) of the KTCs become 2(x1 ⫺ 3) ⫺ u1[⫺3(1 ⫺ x1)2] ⫺ u2 ⫽ 0

(5.21)

2x2 ⫺ u1(⫺1) ⫺ u3 ⫽ 0

(5.22)

u1[(1 ⫺ x1)3 ⫺ x2] ⫽ 0

(5.23)

u2x1 ⫽ 0

(5.24)

u3x2 ⫽ 0

(5.25)

u1, u2, u3 ⭓ 0

(5.26)

At x* ⫽ (1, 0), Eq. (5.21) implies u2 ⫽ ⫺ 4, while to satisfy Eq. (5.24), u2 ⫽ 0. Hence, there exists no Kuhn–Tucker point at the optimum.

Figure 5.1. Feasible region of Example 5.5.

232


Note that when the constraint qualification is violated, it does not necessarily imply that a Kuhn–Tucker point does not exist. To illustrate this, suppose the objective function of Example 5.5 is changed to ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ x22. The optimum still occurs at x* ⫽ (1, 0), and it does not satisfy the constraint qualification. The KTCs given by Eqs. (5.22)–(5.26) will remain the same, while Eq. (5.21) becomes 2(x1 ⫺ 1) ⫺ u1[⫺3(1 ⫺ x1)2] ⫺ u2 ⫽ 0 The reader can easily verify that there exists a Kuhn–Tucker point given by x* ⫽ (1, 0) and u* ⫽ (0, 0, 0) that satisfies the KTCs. The Kuhn–Tucker necessity theorem helps to identify points that are not optimal. In other words, given a feasible point that satisfies the constraint qualification, we can use Theorem 5.1 to prove that it is not optimal if it does not satisfy the KTCs. On the other hand, if it does satisfy the KTCs, there is no assurance that it is optimal to the nonlinear program! For example, consider the following NLP problem. Example 5.6

Minimize

ƒ(x) ⫽ 1 ⫺ x2

Subject to ⫺1 ⭐ x ⭐ 3 Solution. Here, g1(x) ⫽ x ⫹ 1 ⭓ 0 g2(x) ⫽ 3 ⫺ x ⭓ 0 The KTCs are given by ⫺2x ⫺ u1 ⫹ u2 ⫽ 0 ⫺1 ⭐ x ⭐ 3

(5.27) (5.28)

u1(x ⫹ 1) ⫽ 0

(5.29)

u2(3 ⫺ x) ⫽ 0

(5.30)

u 1, u 2 ⭓ 0

(5.31)

Since the constraints are linear, the constraint qualification is satisfied at all feasible points. It is clear that x ⫽ 3 is optimal. But consider the feasible solution x ⫽ 2. To prove that it is not optimal, let us try to construct a Kuhn– Tucker point at x ⫽ 2 that satisfies Eqs. (5.27)–(5.31). To satisfy Eqs. (5.29)

5.5


233

and (5.30), u1 ⫽ u2 ⫽ 0; but x ⫽ 2, u1 ⫽ u2 ⫽ 0 violates Eq. (5.27). Hence, by Theorem 5.1, x ⫽ 2 cannot be optimal. On the other hand, the solution x ⫽ u1 ⫽ u2 ⫽ 0 satisfies Eqs. (5.27)– (5.31) and hence is a Kuhn–Tucker point, but it is not optimal! By Theorem 5.1, we also know that the KTCs must be satisfied at the optimal solution x ⫽ 3. It is easy to verify that the solution x ⫽ 3, u1 ⫽ 0, u2 ⫽ 6 satisfies the KTCs. The following theorem gives conditions under which a Kuhn–Tucker point automatically becomes an optimal solution to the NLP problem. Theorem 5.2

Kuhn–Tucker Sufficiency Theorem

Consider the NLP problem given by Eqs. (5.10)–(5.12). Let the objective function ƒ(x) be convex, the inequality constraints gj (x) be all concave functions for j ⫽ 1, . . . , J, and the equality constraints hk (x) for k ⫽ 1, . . . , K be linear. If there exists a solution (x*, u*, v*) that satisfies the KTCs given by Eqs. (5.13)–(5.17), then x* is an optimal solution to the NLP problem. A rigorous proof of the Kuhn–Tucker sufficiency theorem can be found in Mangasarian [2]. When the sufficiency conditions of Theorem 5.2 hold, finding a Kuhn– Tucker point gives an optimal solution to an NLP problem. Theorem 5.2 can also be used to prove that a given solution to an NLP problem is optimal. To illustrate this, recall Example 5.4: Minimize

ƒ(x) ⫽ x21 ⫺ x2

Subject to g1(x) ⫽ x1 ⫺ 1 ⭓ 0 g2(x) ⫽ 26 ⫺ x21 ⫺ x22 ⭓ 0 h1(x) ⫽ x1 ⫹ x2 ⫺ 6 ⫽ 0 We shall prove that x*1 ⫽ 1, x*2 ⫽ 5 is optimal by using Theorem 5.2. Now, ⵜƒ(x) ⫽ (2x1, ⫺ 1)

and

Hƒ(x) ⫽

冋册 2 0

0 0

Since Hƒ(x) is positive semidefinite for all x, ƒ(x) is a convex function. The inequality constraint g1(x) is linear and hence both convex and concave. To show that g2(x) is concave, compute ⵜg2(x) ⫽ (⫺2x1, ⫺2x2)

and

Hg2(x) ⫽

冋

⫺2

0

0

册

⫺2

234


Since Hg2(x) is negative definite, g2(x) is concave. The equality constraint h1(x) is linear. Hence all the sufficiency conditions of Theorem 5.2 are satisfied, and if we are able to construct a Kuhn–Tucker point using x* ⫽ (1, 5), the solution x* is indeed optimal. The KTCs of Example 5.4 are given below: 2x1 ⫺ u1 ⫹ 2x1u2 ⫺ v1 ⫽ 0

(5.32)

⫺1 ⫹ 2x2u2 ⫺ v1 ⫽ 0

(5.33)

x1 ⫺ 1 ⭓ 0

(5.34)

26 ⫺ x21 ⫺ x22 ⭓ 0

(5.35)

x1 ⫹ x2 ⫺ 6 ⫽ 0

(5.36)

u1(x1 ⫺ 1) ⫽ 0

(5.37)

u2(26 ⫺ x12 ⫺ x22) ⫽ 0

(5.38)

u1, u2 ⭓ 0

(5.39)

Here x* ⫽ (1, 5) satisfies Eqs. (5.34)–(5.36), and hence it is feasible. Equations (5.32) and (5.33) reduce to 2 ⫺ u1 ⫹ 2u2 ⫺ v1 ⫽ 0 ⫺1 ⫹ 10u2 ⫺ v1 ⫽ 0

By setting v1 ⫽ 0, we can get a solution u2 ⫽ 0.1 and u1 ⫽ 2.2. Thus, the solution x* ⫽ (1, 5), u* ⫽ (2.2, 0.1), and v*1 ⫽ 0 satisfies the KTCs. Since the sufficiency conditions of Theorem 5.2 are satisfied, x* ⫽ (1, 5) is an optimal solution to Example 5.4. Note that there also exist other values of u1, u2, v1 that satisfy Eqs. (5.32)–(5.39). Remarks 1. For practical problems, the constraint qualification will generally hold. If the functions are differentiable, a Kuhn–Tucker point is a possible candidate for the optimum. Hence, many of the NLP methods attempt to converge to a Kuhn–Tucker point. (Recall the analogy to the unconstrained optimization case wherein the algorithms attempt to determine a stationary point.) 2. When the sufficiency conditions of Theorem 5.2 hold, a Kuhn–Tucker point automatically becomes the global minimum. Unfortunately, the sufficiency conditions are difficult to verify, and often practical prob-

5.6

SADDLEPOINT CONDITIONS

235

lems may not possess these nice properties. Note that the presence of one nonlinear equality constraint is enough to violate the assumptions of Theorem 5.2 3. The sufficiency conditions of Theorem 5.2 have been generalized further to nonconvex inequality constraints, nonconvex objectives, and nonlinear equality constraints. These use generalizations of convex functions such as quasi-convex and pseudoconvex functions. (See Section 5.9.)

5.6


The discussion of Kuhn–Tucker optimality conditions of Sections 5.4 and 5.5 assume that the objective function and the constraints are differentiable. We now discuss constrained optimality criteria for nondifferentiable functions. Definition A function ƒ(x, y) is said to have a saddlepoint at (x*, y*) if ƒ(x*, y) ⭐ ƒ(x*, y*) ⭐ ƒ(x, y*) for all x and y. The definition of a saddlepoint implies that x* minimizes the function ƒ(x, y*) for all x and y* maximizes the function ƒ(x*, y) for all y. For example, consider a function ƒ(x, y) ⫽ x2 ⫺ xy ⫹ 2y defined over all real values of x and nonnegative values of y. It is easy to verify that the function possesses a saddlepoint at the point x* ⫽ 2, y* ⫽ 4. In other words, ƒ(2, y) ⭐ ƒ(2, 4) ⭐ ƒ(x, 4)

for all y ⭓ 0 and all real x

Recall the Lagrange multiplier method discussed in Section 5.2. It solves a constrained optimization problem of the form Minimize

ƒ(x)


for k ⫽ 1, . . . , K

The Lagrangian function is defined to be L(x; v) ⫽ ƒ(x) ⫺

冘 v h (x) k

k

k

Suppose at v ⫽ v* the minimum of L(x, v*) occurs at x ⫽ x* such that hk (x*) ⫽ 0. We know then, by the Lagrange multiplier method, that x* is an optimal solution to the nonlinear program. It can be shown that (x*, v*) is a saddlepoint of the Lagrangian function satisfying

236


L(x*, v) ⭐ L(x*, v*) ⭐ L(x, v*)

for all x and v

Consider the general NLP problem: Minimize

ƒ(x)


for j ⫽ 1, . . . , J

x僆S The set S may be used to impose additional restrictions on the design variables. For example, the design variables may all be integers or restricted to a certain discrete set. The Kuhn–Tucker saddlepoint problem (KTSP) is as follows: Find (x*, u*) such that L(x*, u) ⭐ L(x*, u*) ⭐ L(x, u*) all u ⭓ 0

all x 僆 S

and

where L(x, u) ⫽ ƒ(x) ⫺

冘 u g (x) j j

j

Theorem 5.3

Sufficient Optimality Theorem

If (x*, u*) is a saddlepoint solution of a KTSP, the x* is an optimal solution to the NLP problem. A proof of this theorem is available in Mangasarian [2, Chap. 3]. Remarks 1. No convexity assumptions of the functions have been made in Theorem 5.3. 2. No constraint qualification is invoked. 3. Nonlinear equality constraints of the form hk (x) ⫽ 0 for k ⫽ 1, . . . , K can be handled easily by redefining the Lagrangian function as L(x, u, v) ⫽ ƒ(x) ⫺

冘 u g (x) ⫺ 冘 v h (x) j j

j

k

k

k

Here the variables vk for k ⫽ 1, . . . , K will be unrestricted in sign. 4. Theorem 5.3 provides only a sufficient condition. There may exist some NLP problems for which a saddlepoint does not exist even though the NLP problem has an optimal solution.

5.6


237

Existence of Saddlepoints. There exist necessary optimality theorems that

guarantee the existence of a saddlepoint solution without the assumption of differentiability. However, they assume that the constraint qualification is met and that the functions are convex. Theorem 5.4

Necessary Optimality Theorem

Let x* minimize ƒ(x) subject to gj (x) ⭓ 0, j ⫽ 1, . . . , J and x 僆 S. Assume S is a convex set, ƒ(x) is a convex function, and gj (x) are concave functions on S. Assume also that there exists a point x 僆 S such that gj (x) ⬎ 0 for all j ⫽ 1, 2, . . . , J. then there exists a vector of multipliers u* ⭓ 0 such that (x*, u*) is a saddlepoint of the Lagrangian function L(x, u) ⫽ ƒ(x) ⫺

冘 u g (x) j j

j

satisfying L(x*, u) ⭐ L(x*, u*) ⭐ L(x, u*) for all x 僆 S and u ⭓ 0. For a proof of this theorem, refer to the text by Lasdon [3, Chap. 1]. Even though Theorem 5.3 and the KTSP provide sufficient conditions for optimality without invoking differentiability and convexity, determination of a saddlepoint to a KTSP is generally difficult. However, the following theorem makes it computationally more attractive. Theorem 5.5 A solution (x*, u*) with u* ⭓ 0 and x* 僆 S is a saddlepoint of a KTSP if and only if the following conditions are satisfied: (i) x* minimizes L(x, u*) over all x 僆 S (ii) gj (x*) ⭓ 0 for j ⫽ 1, . . . , J (iii) ujgj (x*) ⫽ 0 for j ⫽ 1, . . . , J For a proof, see Lasdon [3, Chap. 1]. Condition (i) of Theorem 5.5 amounts to finding an unconstrained minimum of a function, and any of the direct-search methods discussed in Chapter 3 could be used. Of course, this assumes prior knowledge of the value of u*. However, a trial-and-error method can be used to determine u* and x* simultaneously and also satisfy conditions (ii) and (iii). One such method, due to Everett [4], is called the generalized Lagrange multiplier method (Section 5.8). Theorems 5.3 and 5.5 also form the basis of many of the Lagrangian

238


relaxation methods that have been developed for solving large-scale NLP problems [3]. It is important to note that saddlepoints may not exist for all NLP problems. The existence of saddlepoints is guaranteed only for NLP problems that satisfy the conditions of Theorem 5.4.

5.7

SECOND-ORDER OPTIMALITY CONDITIONS

In Sections 5.4–5.6, we discussed the first-order necessary and sufficient conditions, called the Kuhn–Tucker conditions, for constrained optimization problems using the gradients of the objective function and constraints. Second-order necessary and sufficient optimality conditions that apply to twicedifferentiable functions have been developed by McCormick [5], whose main results are summarized in this section. Consider the following NLP problem. Problem P1 Minimize

ƒ(x)


j ⫽ 1, 2, . . . , J

hk (x) ⫽ 0

k ⫽ 1, 2, . . . , K

x 僆 RN The first-order KTCs are given by ⵜƒ(x) ⫺

冘 u ⵜg (x) ⫺ 冘v ⵜh (x) ⫽ 0 j

j

k

k

(5.40)

j

gj (x) ⭓ 0

j ⫽ 1, . . . , J

(5.41)

hk (x) ⫽ 0

k ⫽ 1, . . . , K

(5.42)

ujgj (x) ⫽ 0

j ⫽ 1, . . . , J

(5.43)

uj ⭓ 0

j ⫽ 1, . . . , J

(5.44)

Definitions x is a feasible solution to an NLP problem when gj(x) ⭓ 0 for all j and hk(x) ⫽ 0 for all k. x* is a local minimum to an NLP problem when x* is feasible and ƒ(x*) ⭐ ƒ(x) for all feasible x in some small neighborhood ␦(x*) of x*. x* is a strict (unique or isolated) local minimum when x* is feasible and ƒ(x*) ⬍ ƒ(x) for feasible x ⫽ x* in some small neighborhood ␦(x*) of x*.

5.7


239

A Kuhn–Tucker point to an NLP problem is a vector (x*, u*, v*) satisfying Eqs. (5.40)–(5.44). Let us first consider the basic motivation for the second-order optimality conditions. For simplicity, consider an equality-constrained NLP problem as follows: Minimize

ƒ(x)


k ⫽ 1, 2, . . . , K

The first-order KTCs are given by hk (x) ⫽ 0, k ⫽ 1, . . . , k, ⵜƒ(x) ⫺

冘 v ⵜh (x) ⫽ 0 k

(5.45)

k

k

Consider a point x that satisfied the first-order conditions. To check further whether it is a local minimum, we can write down the Taylor series expansion at the point x using higher order terms for each function ƒ and hk as follows: ⵜƒ(x) ⫽ ƒ(x ⫹ ⌬x) ⫺ ƒ(x) ⫽ ⵜƒ(x) ⌬x ⫹ –12 ⌬xT Hƒ ⌬x ⫹ O(⌬x)

(5.46)

where O(⌬x) are very small higher order terms involving ⌬x. ⌬hk (x) ⫽ hk (x ⫹ ⌬x) ⫺ hk (x) ⫽ ⵜhk (x) ⌬x ⫹ –12 ⌬xT Hk ⌬x ⫹ O(⌬x)

(5.47)

where Hk is the Hessian matrix of hk (x) evaluated at x. Multiply Eq. (5.47) by the Kuhn–Tucker multiplier vk, and sum over all k ⫽ 1, . . . , K. Subtracting this sum from Eq. (5.46), we obtain ⌬ƒ(x) ⫺

冘 v ⌬h (x) ⫽ 冋ⵜƒ(x) ⫺ 冘 v ⵜh (x)册 ⌬x ⫹ – ⌬x 冋H ⫺ 冘 v H 册 ⌬x ⫹ O(⌬x) k

k

k

k

k

k

1 2

T

ƒ

k

k

(5.48)

k

For x ⫹ ⌬x to be feasible, ⌬hk (x) ⫽ 0

(5.49)

Assuming that the constraint qualification is satisfied at x, the Kuhn–Tucker necessary theorem implies that

240


ⵜƒ(x) ⫺

冘 v ⵜh (x) ⫽ 0 k

k

(5.50)

k

Using Eqs. (5.49) and (5.50), Eq. (5.48) reduces to

冋

⌬ƒ(x) ⫽ –12 ⌬xT Hƒ ⫺

冘 v H 册 ⌬x ⫹ O(⌬x) k

k

(5.51)

k

For x to be a local minimum, it is necessary that ⌬ƒ(x) ⭓ 0 for all feasible movement ⌬x around x. Using Eqs. (5.49) and (5.51), the above condition implies that

冋

⌬xT Hƒ ⫺

冘 v H 册 ⌬x ⬎ 0 k

k

(5.52)

k

for all ⌬x satisfying ⌬hk (x) ⫽ 0

for k ⫽ 1, . . . , K

(5.53)

Using Eq. (5.47) and ignoring the second and higher order terms in ⌬x, Eq. (5.53) reduces to ⌬hk (x) ⫽ ⵜhk (x) ⌬x ⫽ 0

Thus assuming that the constraint qualification is satisfied at x, the necessary conditions for x to be a local minimum are as follows: 1. There exists vk, k ⫽ 1, . . . , K, such that (x, v) is a Kuhn–Tucker point. 2. ⌬xT[Hƒ ⫺ vkHk] ⌬x ⭓ 0 for all ⌬x satisfying

冘 k

ⵜhk (x) ⌬x ⫽ 0

for k ⫽ 1, . . . , K

Similarly, the sufficient condition for x to be strict local minimum is given by ⌬ƒ(x) ⬎ 0

for all feasible ⌬x around x

This implies that

冋

⌬xT Hƒ ⫺

for all ⌬x satisfying

冘 v H 册 ⌬x ⬎ 0 k

k

k

5.7


ⵜhk (x) ⌬x ⫽ 0

for all k ⫽ 1, . . . , K

241

(5.54)

We shall now present the formal statements of second-order necessary and sufficient conditions for an NLP problem involving both equality and inequality constraints. Theorem 5.6

Second-Order Necessity Theorem

Consider the NLP problem given by Problem P1. Let ƒ, g, and h be twicedifferentiable functions, and let x* be feasible for the nonlinear program. Let the active constraint set at x* be I ⫽ {j兩gj (x*) ⫽ 0}. Furthermore, assume that ⵜgj (x*) for j 僆 I and ⵜhk (x*) for k ⫽ 1, 2, . . . , K are linearly independent. Then the necessary conditions that x* be a local minimum to the NLP problem are as follows: 1. There exists (u*, v*) such that (x*, u*, v*) is a Kuhn–Tucker point. 2. For every vector y(1⫻N) satisfying ⵜgj (x*)y ⫽ 0

for j 僆 I

(5.55)

ⵜhk (x*)y ⫽ 0

for k ⫽ 1, 2, . . . , K

(5.56)

it follows that yTHL(x*, u*, v*)y ⭓ 0

(5.57)

where

冘 u g (x) ⫺ 冘 v h (x) J

L(x, u, v) ⫽ ƒ(x) ⫺

K

j j

j⫽1

k

k

k⫽1

and HL(x*, u*, v*) is the Hessian matrix of the second partial derivatives of L with respect to x evaluated at (x*, u*, v*). We shall illustrate Theorem 5.6 with an example in which the first-order necessary conditions are satisfied while the second-order conditions show that the point is not optimal. Example 5.7 [5]

Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ x22

Subject to g1(x) ⫽ ⫺x1 ⫹ x22 ⭓ 0 Suppose we want to verify whether x* ⫽ (0, 0) is optimal.

242


Solution ⵜƒ(x) ⫽ [2(x1 ⫺ 1), 2x2] ⵜg1(x) ⫽ (⫺1, 2x2)

I ⫽ {1}

Since ⵜg1(x*) ⫽ (⫺1, 0) is linearly independent, the constraint qualification is satisfied at x*. The first-order KTCs are given by 2(x1 ⫺ 1) ⫹ u1 ⫽ 0 2x2 ⫺ 2x2u1 ⫽ 0 u1(⫺x1 ⫹ x22) ⫽ 0 u1 ⭓ 0 Here x* ⫽ (0, 0) and u*1 ⫽ 2 satisfy the above conditions. Hence, (x*, u*) ⫽ (0, 0, 2) is Kuhn–Tucker point and x* satisfies the first-order necessary conditions of optimality by Theorem 5.1. In other words, we do not know whether or not (0, 0) is an optimal solution to the NLP problem! Let us now apply the second-order necessary conditions to test whether (0, 0) is a local minimum to the NLP problem. The first part of Theorem 5.6 is already satisfied, since (x*, u*) ⫽ (0, 0, 2) is a Kuhn–Tucker point. To prove the second-order conditions, compute HL(x, u) ⫽

冋

册

2 0 0 2 ⫺ 2u1

At (x*, u*), HL(x*, u*) ⫽

冋册 2 0

0

⫺4

We have to verify whether

冋册

yT

2 0

0

⫺4

y⭓0

for all y satisfying ⵜg1(x*)y ⫽ 0

or

冉冊

(⫺1, 0)

y1 ⫽0 y2

In other words, we need to consider only vectors (y1, y2) of the form (0, y2) to satisfy Eq. (5.57). Now,

5.7


冋册冋册

(0, y2)

2 0

0

⫺4

0 ⫽ ⫺4y22 ⬍ 0 y2

243

for all y2 ⫽ 0

Thus, x* ⫽ (0, 0) does not satisfy the second-order necessary conditions, and hence its is not a local minimum for the NLP problem. Sufficient Conditions. When a point satisfies the second-order necessary con-

ditions given by Theorem 5.6, it becomes a Kuhn–Tucker point and a candidate for a local minimum. To show that it is in fact a minimum point, we need the second-order sufficient conditions. Of course, when a nonlinear program satisfies the assumptions of Theorem 5.2 (Kuhn–Tucker sufficiency theorem), the Kuhn–Tucker point automatically becomes the global minimum. However, Theorem 5.2 requires that the objective function be convex, the inequality constraints concave, and the equality constraints linear. These assumptions are too rigid and may not be satisfied in practice very often. In such situations, the second-order sufficiency conditions may be helpful in showing that a Kuhn–Tucker point is a local minimum. Theorem 5.7

Second-Order Sufficiency Theorem

Sufficient conditions that a point x* is a strict local minimum of the NLP problem P1, where ƒ, gj, and hk are twice-differentiable functions, are as follows: (i) There exists (u*, v*) such that (x*, u*, v*) is a Kuhn–Tucker point. (ii) For every nonzero vector y(1⫻N) satisfying ⵜgj (x*)y ⫽ 0

j 僆 I1 ⫽ {j兩gj (x*) ⫽ 0, u* j ⬎ 0}

(5.58)

ⵜgj (x*)y ⭓ 0

j 僆 I2 ⫽ {j兩gj (x*) ⫽ 0, u* j ⫽ 0}

(5.59)

ⵜhk (x*)y ⫽ 0

k ⫽ 1, 2, . . . , K

(5.60)

y ⫽0 it follows that yTHL(x*, u*, v*)y ⬎ 0

(5.61)

Note: I1 傼 I2 ⫽ I, the set of all active constraints at x*. Comparing Theorems 5.6 and 5.7, it is clear that the sufficient conditions add very few new restrictions to the necessary conditions, and no additional assumptions about the properties of the functions are needed. The minor

244


changes are that Eq. (5.55) need not be satisfied for all active constraints and inequality (5.57) has to be satisfied as a strict inequality. Remark The restrictions on vector y given by Eqs. (5.58) and (5.59) use information on the multiplier u*. Han and Mangasarain [6] have given an equivalent form without the use of the Kuhn–Tucker multiplier u*. They prove that Eqs. (5.58) and (5.59) are equivalent to the following set of conditions: ⵜƒ(x*)y ⭐ 0 ⵜgj (x*)y ⭓ 0

for j 僆 I ⫽ {j兩gj (x*) ⫽ 0}

We now illustrate the use of the second-order sufficient conditions. Example 5.8

Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ x22

Subject to g1(x) ⫽ ⫺x1 ⫹ –51 x22 ⭓ 0 This problem is very similar to the one given in Example 5.7. Suppose we want to verify whether x* ⫽ (0, 0) is a local minimum. Note that the region S ⫽ {x兩g1(x) ⭓ 0} is not a convex set. For example, x ⫽ (0.2, 1) and (0.2, ⫺1) are feasible, but the midpoint (0.2, 0) is not. The KTCs for this problem are 2(x1 ⫺ 1) ⫹ u1 ⫽ 0 2x2 ⫺ –25 x2u1 ⫽ 0 u1(⫺x1 ⫹ –15 x22) ⫽ 0 u1 ⭓ 0 Here x* ⫽ (0, 0), u* 1 ⫽ 2 satisfies the KTCs. Using Theorem 5.1, we can conclude that x* ⫽ (0, 0) satisfies the necessary conditions for the minimum. But we cannot conclude that x* is a local minimum, since the function g1(x) is convex and violates the assumptions of Theorem 5.2 (Kuhn–Tucker sufficiency theorem). Using the second-order sufficient conditions, we find that

5.8

GENERALIZED LAGRANGE MULTIPLIER METHOD

HL(x*, u*) ⫽

245

冋册 2 0

0 1.2

The vector y ⫽ (y1, y2) satisfying Eqs. (5.58) and (5.59) is of the form (0, y2) as in Example 5.8. Inequality (5.61) reduces to

冋册冋册

(0, y2)

2 0 0 1.2

0 ⫽ 1.2y22 ⬎ 0 y2

for all y2 ⫽ 0

Hence by Theorem 5.7, x* ⫽ (0, 0) is a strict local minimum. Fiacco [7] has extended Theorem 5.7 to sufficient conditions for weak (not necessarily strict) minimum also.

5.8


The usefulness of the Lagrange multiplier method for solving constrained optimization problems is not limited to differentiable functions. Many engineering problems may involve discontinuous or nondifferentiable functions to be optimized. Everett [4] generalized the Lagrange multiplier method presented earlier to handle such problems. Consider the NLP problem Minimize

ƒ(x)

Subject to gj (x) ⭓ bj

for j ⫽ 1, 2, . . . , J

x僆S where S is a subset of RN, imposing additional restrictions on the variables xj (e.g., S may be a discrete set). Everett’s generalized Lagrangian function corresponding to the NLP problem is given by

冘 ␭ g (x) J

E(x; ␭) ⫽ ƒ(x) ⫺

j j

(5.62)

j⫽1

where the ␭j’s are nonnegative multipliers. Suppose the unconstrained minimum of E(x; ␭) over all x 僆 S is attained at the point x for a fixed value of ␭ ⫽ ␭. Then, Everett [4] proved that x is an optimal solution to the following mathematical program:

246


Minimize

ƒ(x)

Subject to gj (x) ⭓ gj (x)

j ⫽ 1, . . . , J

x僆S Hence, to solve the original NLP problem, it is sufficient to find nonnegative multipliers ␭* (called Everett’s multipliers) such that the unconstrained minimum of E(x; ␭*) over all x 僆 S occurs at the point x* such that gj (x*) ⫽ bj

for j ⫽ 1, . . . , J

(5.63)

We call this Everett’s condition. Since any of the search methods discussed in Chapter 3 could be used to minimize Eq. (5.62), Everett’s method looks computationally attractive. However, for many practical problems, Everett’s condition is too restrictive since Eq. (5.63) implies that all the constraints have to be active at the optimum. Hence, Everett indicates that if the constraints gj (x*) are ‘‘close to’’ the righthand-side constants bj, then x* is a good approximate optimal solution to the given NLP problem. Thus, the basic steps of Everett’s method to find an approximate optimal solution to the NLP problem are as follows: Step 1. Choose an initial set of nonnegative multipliers ␭(1) for j ⫽ 1, . . . , j J. Step 2. Find the unconstrained minimum of E(x; ␭(1)) over x 僆 S by any direct-search method. Let the constrained minimum occur at x ⫽ x(1). Step 3. Compute gj(x(1)) for j ⫽ 1, . . . , J. If the values of gj(x(1)) are close to bj for all j (e.g., within specified error tolerance), then terminate. Otherwise, ‘‘update’’ the multipliers to a new set of values ␭(2) and repeat the optimization process. j Everett also provided a rule for updating the multipliers systematically in step 3. Recall that the multipliers ␭’s can be interpreted as shadow prices. In other words, if gj (x) ⭓ bj corresponds to the minimum level of production for product j, then ␭j can be interpreted as the ‘‘price’’ of product j to break even. Hence, if ␭j is increased keeping all the other shadow prices fixed, you would expect to produce and sell more of product j. This is the basis of Everett’s theorem, which shows that gj (x) monotonically increases with ␭j while holding all other ␭’s fixed. Theorem 5.8 Let ␭(1) and ␭(2) be nonnegative vectors such that ␭(2) ⬎ ␭i(1) i

and

␭j(1) ⫽ ␭j(2)

for all j ⫽ i

If x(1) and x(2) minimize E(x; ␭) given by Eq. (5.62), then

5.8


247

gi[x(1)] ⭓ gi[x(2)] We shall illustrate Everett’s basic algorithm with the following example: Example 5.9

Minimize

ƒ(x) ⫽ x21 ⫹ x22

Subject to g1(x) ⫽ 2x1 ⫹ x2 ⭓ 2 Everett’s function is given by E(x; ␭) ⫽ x21 ⫹ x22 ⫺ ␭(2x1 ⫹ x2) We begin Everett’s method with ␭ ⫽ ␭1 ⫽ 0. The unconstrained minimum of E(x; 0) occurs at the point x(1) ⫽ (0, 0). Since g1(x(1)) ⫽ 0, which is less than 2, we increase ␭ to increase g1(x). Choose ␭ ⫽ ␭2 ⫽ 1. The unconstrained minimum of E(x; 1) occurs at the point x(2) ⫽ (1, –12 ), g(x(2)) ⫽ 2 ⫹ –12 ⬎ 2. Hence, ␭ has to be decreased to get a solution that reduces the constraint value. The remaining steps are shown in Table 5.1. Note that the value of ␭ in each step is simply the midpoint of the two previous ␭’s, since we know that the optimal ␭ is between 0 and 1. The convergence is achieved at step 8. When several constraints are present, the adjustment of ␭’s becomes more difficult since Theorem 5.8 does not hold when several ␭’s are changed simultaneously. Even if we decide to change one ␭j at a time, it is quite possible that it may affect other constraint values. This is especially true if there are a lot of dependencies among the constraints. On the other hand, if the constraints are relatively ‘‘independent,’’ then Everett’s method may eventually converge.

Table 5.1

Everett’s Method for Example 5.9

Step t

␭t

1 2 3 4 5 6 7 8

0 1 0.5 0.75 0.88 0.82 0.78 0.8

(t) x(t) ⫽ (x(t) 1 , x2 )

x(1) x(2) x(3) x(4) x(5) x(6) x(7) x(8)

⫽ ⫽ ⫽ ⫽ ⫽ ⫽ ⫽ ⫽

(0, 0) (1, 0.5) (0.5, 0.25) (0.75, 0.375) (0.88, 0.44) (0.82, 0.41) (0.78, 0.39) (0.8, 0.4)

g1(x(t)) 0 2.5 1.25 1.875 2.2 2.05 1.95 2

Constraint Violation ⬍2 ⬎2 ⬍2 ⬍2 ⬎2 ⬎2 ⬍2

2

248


Improvements to Everett’s Method. Brooks and Geoffrion [8] observed that when there are inactive constraints in the NLP problem, Everett’s conditions are unnecessarily restrictive because when a multiplier is zero, it is not necessary to require that the corresponding constraint be satisfied as a strict equality (or an approximate equality). Thus, they weakened Everett’s conditions such that when their conditions are satisfied, the global optimum is guaranteed. Modified Everett’s Conditions

1. x* minimizes E(x; ␭*) given by Eq. (5.62) over all x 僆 S, and ␭* ⭓ 0. 2. When ␭*j ⬎ 0: gj (x*) ⫽ bj ␭j* ⫽ 0: gj (x*) ⭓ bj for j ⫽ 1, . . . , J If the above conditions are satisfied, then x* is optimal to the NLP problem: Minimize

ƒ(x)

Subject to gj (x) ⭓ bj

j ⫽ 1, . . . , J

x僆S From Theorem 5.5, it is clear that the modified Everett’s conditions are nothing more than the requirement that (x*, ␭*) is a saddlepoint of the Lagrangian function:

冘 ␭ [g (x) ⫺ b ] J

L(x, ␭) ⫽ ƒ(x) ⫺

j

j

j

j⫽1

Thus, with the modification by Brooks and Geoffrion, Everett’s method becomes an iterative technique for constructing a saddlepoint to the KTSP. Of course, the existence of a saddlepoint is guaranteed only under the assumptions given in Theorem 5.4. In addition to weakening Everett’s optimality conditions, Brooks and Geoffrion also gave a method for updating the multipliers required in step 3 of Everett’s method. They observed that when we have an LP problem (i.e., when S is the nonnegative orthant, ƒ and gj are linear functions), a solution (x*, ␭*) satisfies the modified Everett’s conditions if and only if x* is optimal for the given linear program and ␭* is optimal for its dual. We know from the results of Chapter 4 that the optimal dual variables are nothing but the shadow prices of the primal constraints. We also know that the optimal Lagrange multipliers correspond to the shadow prices in the nonlinear program as well. Hence, an approximation to Everett’s multipliers (␭j) can be obtained with a linear approximation of the nonlinear program.

5.9

GENERALIZATION OF CONVEX FUNCTIONS

249

Brooks and Geoffrion’s Updating Scheme. Suppose we have the solutions x(1), x(2), . . . , x(t⫺1) corresponding to the unconstrained minimum of E(x; ␭) at ␭ ⫽ ␭(1), ␭(2), . . . , ␭(t⫺1). To determine the updated values of the multipliers ␭(t) for use in step t, we solve the following linear program in variables yi:

冘 ƒ(x )y 冘y ⫽1 冘 g (x )y ⭓ b

t⫺1

(i)

Minimize

i

i⫽1 t⫺1

Subject to

i

i⫽1 t⫺1

(i)

j

i

j

for j ⫽ 1, . . . , J

i⫽1

yi ⭓ 0 The optimal value of the dual variables corresponding to the inequality constraints (j ⫽ 1, . . . , J) will become the new multipliers ␭(t) j needed for step t. Note from the duality theory of linear programming that the dual variables ␭(t) will automatically be nonnegative. Since we are only interested in the j optimal dual solution to the above LP problem (and are not concerned with the optimal yi values), one could simply solve the dual LP problem directly:

冘 b␭ ␭ ⫹ 冘 g (x )␭ ⭐ ƒ(x ) J

Maximize ␭0 ⫹

j j

j⫽1 J

(i)

0

j

(i)

j

for i ⫽ 1, 2, . . . , t ⫺ 1.

j⫽1

␭j ⭓ 0

for j ⫽ 1, . . . , J

␭0 unrestricted in sign

Note that the dual LP problem to be solved at step t ⫹ 1 in order to determine ␭(t⫹1) would have just one additional constraint. Hence, the dual simplex method described in Chapter 4 can be used to determine the new dual optimal solution without re-solving the problem. Of course, the algorithm is terminated when we find a ␭* such that (x*, ␭*) satisfies the modified Everett’s conditions.

5.9


In Section 5.5 we developed the KTCs for NLP problems. When the sufficiency conditions of Theorem 5.2 hold, a Kuhn–Tucker point automatically

250


becomes a global minimum to the NLP problem. The sufficiency conditions required that ‘‘greater than or equal to’’ type constraints are concave functions, equality constraints are linear, and the objective function to minimize is convex. Note that the presence of one nonlinear equality constraint is enough to violate the sufficiency conditions. The sufficiency conditions of Theorem 5.2 have been generalized further to nonconcave inequality constraints, nonconvex objective functions, and nonlinear equality constraints. They use generalizations of convex functions such as pseudoconvex and quasi-convex functions. Definition: Pseudoconvex Function A differentiable function ƒ(x) defined on an open convex set S is pseudoconvex on S if and only if for all x (1), x (2) 僆 S ⵜƒ(x (1))(x (2) ⫺ x (1)) ⭓ 0 ⇒ ƒ(x (2)) ⭓ ƒ(x (1))

Remarks 1. ƒ(x) is pseudoconcave if ⫺ƒ(x) is pseudoconvex. 2. Every convex function is also pseudoconvex, but a pseudoconvex function may not be convex. Figures 5.2 and 5.3 illustrate pseudoconvex and pseudoconcave functions. Definition: Strictly Quasi-Convex Function A function ƒ(x) defined on a convex set S is strictly quasi-convex on S if and only if

Figure 5.2. Pseudoconvex function.

5.9


251

Figure 5.3. Pseudoconcave function.

ƒ(␭x (1) ⫹ (1 ⫺ ␭)x (2)) ⬍ max[ƒ(x (1)), ƒ(x (2))] for all x (1), x (2) 僆 S

0⬍␭⬍1

ƒ(x (1)) ⫽ ƒ(x (2))

Definition: Quasi-Convex Function A function ƒ(x) defined on a convex set S is quasi-convex on S if and only if ƒ(␭x (1) ⫹ (1 ⫺ ␭)x (2)) ⭐ max[ƒ(x (1)), ƒ(x (2))] for all x (1), x (2) 僆 S

0⭐␭⭐1

Figures 5.4 and 5.5 illustrate quasi-convex and strictly quasi-convex functions.

Figure 5.4. Quasi-convex function.

252


Figure 5.5. Strictly quasi-convex function.

Remarks 1. A pseudoconvex function is also a quasi-convex function. But a quasiconvex function may not be pseudoconvex. 2. A strictly quasi-convex function need not necessarily be quasi-convex unless ƒ(x) is assumed to be continuous on the convex set of S. Example 5.10

冦

x ƒ(x) ⫽ 0 x⫺1

for x ⭐ 0 for 0 ⭐ x ⭐ 1 for x ⭓ 1

The above function is both quasi-convex and quasi-concave but neither strictly quasi-convex nor strictly quasi-concave. Example 5.11

ƒ(x) ⫽ x3

for x 僆⺢

The above function is both strictly quasi-convex and strictly quasi-concave, but it is neither quasi-convex nor quasi-concave because of the inflection point at x ⫽ 0. Theorem 5.9 Let ƒ be pseudoconvex function defined on an (open) convex set S. If ⵜƒ(x)

⫽ 0, then x minimizes ƒ(x) over all x 僆 S.

5.9


253

Theorem 5.10 Let ƒ be a strictly quasi-convex function defined on a convex set S. Then a local minimum of ƒ is also a global minimum. Remarks 1. For a proof of Theorems 5.9 and 5.10, the reader is referred to Mangasarian [2]. 2. Theorem 5.9 need not be true for strictly quasi-convex and quasi-convex functions. 3. Theorem 5.10 holds for pseudoconvex functions also. 4. Theorem 5.10 need not hold for quasiconvex functions. 5. We can show that every nonnegative linear combination of convex functions is also a convex function. But this is not true in general for pseudoconvex, strictly quasi-convex, and quasi-convex functions. Theorem 5.11 Generalization of Kuhn–Tucker Sufficient Optimality Theorem Consider the NLP problem Minimize

ƒ(x)

Subject to gj (x) ⭓ 0 hk (x) ⫽ 0

for j ⫽ 1, 2, . . . , J for k ⫽ 1, 2, . . . , K

x ⫽ (x1, x2, . . . , xN) KTP is as follows: Find x, u, and v such that ⵜƒ(x) ⫺

冘 u ⵜg (x) ⫺ 冘 v ⵜh (x) ⫽ 0 j

j

j

uj ⭓ 0

k

k

k

vk unrestricted in sign

ujgj (x) ⫽ 0 gj (x) ⭓ 0

for all j hk (x) ⫽ 0

Let ƒ(x) be pseudoconvex, gj be quasi-concave, and hk be both quasi-convex and quasi-concave. If (x, u, v) solves the KTP, then x solves the NLP problem. For a proof of Theorem 5.11, the reader is referred to Mangasarian [2] and Bazaraa et al. [1].

254

5.10


SUMMARY

In this chapter we developed necessary and sufficient conditions of optimality for constrained optimization problems. We began with the discussion of Lagrangian optimality conditions for problems with equality constraints. These were then extended to inequality constraints in the form of Kuhn–Tucker optimality conditions, which are first-order conditions involving the gradients of the objective function and the constraints. We learned that the Kuhn–Tucker conditions are necessary when the functions are differentiable and the constraints satisfy some regularity condition known as the constraint qualification. Kuhn–Tucker conditions become sufficient conditions for global minima when the objective function is convex, the ‘‘inequality’’ constraints are concave functions, and the equality constraints are linear. We also discussed saddlepoint optimality conditions that could be applicable if the functions were not differentiable. Since there could be several points satisfying the Kuhn–Tucker necessary conditions, we developed second-order necessary conditions that must be satisfied for a point to be a local minimum. Similarly, the assumptions under which the Kuhn–Tucker sufficiency conditions hold are quite rigid. Hence, second-order sufficiency conditions were developed that do not require convexity of the functions and linearity of the equality constraints. Both the necessary and sufficient second-order conditions impose additional restrictions over and above those given by Kuhn–Tucker and hence can be useful in reducing the set of candidate optima.

REFERENCES 1. Bazaraa, M. S., D. Sherali, and C. M. Shetty, Nonlinear Programming: Theory and Algorithms, 2nd ed., Wiley, New York, 1993. 2. Mangasarian, O. L., Nonlinear Programming, McGraw-Hill, New York, 1969. 3. Lasdon, L. S., Optimization Theory for Large Systems, Macmillan, New York, 1979. 4. Everett, H., ‘‘Generalized Lagrange Multiplier Method for Solving Problems of Optimum Allocation of Resources,’’ Oper. Res., 11, 399–471 (1963). 5. McCormick, G. P., ‘‘Second Order Conditions for Constrained Optima,’’ SIAM J. Appl. Math., 15, 641–652 (1967). 6. Han, S. P., and O. L. Mangasarian, ‘‘Exact Penalty Functions in Nonlinear Programming,’’ Math. Programming, 17, 251–269 (1979). 7. Fiacco, A. V., ‘‘Second Order Sufficient Conditions for Weak and Strict Constrained Minima,’’ SIAM J. Appl. Math., 16, 105–108 (1968). 8. Brooks, R., and A. Geoffrion, ‘‘Finding Everett’s Lagrange Multipliers by Linear Programming,’’ Oper. Res., 14, 1149–1153 (1966).

PROBLEMS

255

PROBLEMS

5.1. Explain the difficulties that will be encountered if Lagrange multipliers are used with problems involving nonnegative variables. 5.2. What is the significance of the constraint qualification? 5.3. What is a saddlepoint solution? What is its significance in constrained optimization? 5.4. Under what conditions do saddlepoint solutions exist for NLP problems? 5.5. What are the primary uses of second-order necessary and sufficient conditions? 5.6. Consider the minimization of ƒ(x) over a feasible region denoted by S. Suppose ⵜƒ(x) ⫽ 0 for every x 僆 S. What can you say about the nature of optimal solution to this problem and why? 5.7. Use the Lagrange multiplier method to solve Minimize

z ⫽ x21 ⫹ x22 ⫹ x23

Subject to x1 ⫹ 2x2 ⫹ 3x3 ⫽ 7 2x1 ⫹ 2x2 ⫹ x3 ⫽ –29 Explain (in two or three lines) why your procedure guarantees an optimum solution. 5.8. Use the Lagrange multiplier method to determine the global minimum and the global maximum of the problem ƒ(x) ⫽ x2 ⫺ x21 Subject to x12 ⫹ x22 ⫽ 1

with x1, x2 unrestricted in sign

5.9. Find the shortest distance from the point (1, 0) to the parabola y2 ⫽ 4x by: (a) Eliminating the variable y (b) The Lagrange multiplier technique Explain why procedure (a) fails to solve the problem, while procedure (b) does not fail. 5.10. Given the problem

256


Minimize

ƒ(x) ⫽ (x1 ⫺ 3)2 ⫹ (x2 ⫹ 4)2 ⫹ e5x3

Subject to x1 ⫹ x2 ⫹ x3 ⭐ 1 x1, x2, x3 ⭓ 0 (a) Write the Kuhn–Tucker conditions for this problem. (b) Show that Kuhn–Tucker conditions are sufficient for the existence of an optimum solution to the given problem. (c) Prove, using (a) and (b), that x ⫽ (1, 0, 0) is optimal. 5.11. Consider the nonlinear program Minimize

Z ⫽ x21 ⫹ 2x22

Subject to x21 ⫹ x22 ⭐ 5 2x1 ⫺ 2x2 ⫽ 1 (a) Write the equivalent Kuhn–Tucker problem. (b) Using (a), what can you conclude about the following solutions to the nonlinear program? (i) x (1) ⫽ (0, 0) (ii) x (2) ⫽ (1, –12 ) (iii) x (3) ⫽ (–13 , ⫺ –16 ) You must justify your conclusions. Quote any theorem that is applicable to your conclusions. 5.12. Consider the NLP problem given by Minimize

ƒ(x) ⫽ 1 ⫺ 2x1 ⫺ 4x1x2

Subject to x1 ⫹ 4x1x2 ⭐ 4 x1 ⭓ 0

x2 ⭓ 0

(a) Show that the feasible region of the above NLP problem is not a convex set. (b) Write down the Kuhn–Tucker conditions for this NLP problem. (c) Now let y1 ⫽ x1 and y2 ⫽ x1x2. Write the given problem in terms of (y1, y2). What kind of mathematical program is this? Show that the feasible region of the transformed problem is a convex set. Solve the transformed problem for optimal values of y, say (y1, y2). Obtain the corresponding values (x1, x2). (d) Prove that the constraint qualification is satisfied at x. (You must show this analytically.)

PROBLEMS

257

(e) Prove that x is a candidate optimum using the Kuhn–Tucker necessity theorem. (f) Are the Kuhn–Tucker conditions sufficient to show x is an optimal solution to the NLP problem? Why or why not? Explain. 5.13. Consider the nonlinear programming problem Maximize

z ⫽ a1x1 ⫹ x2

Subject to x21 ⫹ x22 ⭐ 25 x1 ⫺ x2 ⭐ 1 x1 ⭓ 0

x2 ⭓ 0

(a) Write down the Kuhn–Tucker conditions. (b) Show that the Kuhn–Tucker conditions are both necessary and sufficient. (c) Using (b), determine the range of values of a1 for which the solution x1 ⫽ 4, x2 ⫽ 3 is optimal. 5.14. (a) What are the practical uses of the Kuhn–Tucker (i) necessity theorem and (ii) sufficiency theorem? State under what assumptions they are applicable. (b) Write down the Kuhn–Tucker conditions for the following problem. Simplify and give the final form. Maximize

z ⫽ 3x21 ⫺ 2x2

Subject to

2x1 ⫹ x 2 ⫽ 4 x12 ⫹ x22 ⭐ 40 x1 ⭓ 0 x2 unrestricted in sign

5.15. Consider the NLP problem

冘x N

Maximize

z⫽

j

j⫽1

Subject to

冘 Kx ⭐ D N

2 j

j⫽1

j

xj ⭓ 0

for all j

where the Kj’s and D are positive constants.

258


(a) Write down the Kuhn–Tucker conditions. (b) Show that a solution to the Kuhn–Tucker conditions gives an optimal solution to the NLP problem. (c) Using (b), find an optimal solution to the NLP problem. 5.16. Consider the problem Minimize

ƒ(x) ⫽ x21 ⫹ x22

Subject to

(x1 ⫺ 1)3 ⫺ x22 ⫽ 0

(a) Show graphically that the optimal solution is at x1 ⫽ 1, x2 ⫽ 0. (b) Apply the Lagrange multiplier method to solve this problem. Can you explain why it fails? 5.17. Consider the NLP problem given in problem 5.12. Verify whether or not the second-order necessary and sufficient conditions are satisfied at the candidate optimum x. 5.18. Consider the following NLP problem: Minimize

ƒ(x) ⫽ 0.2x31 / 2 ⫹ 0.05x22 ⫺ 3x1 ⫺ 5x2

Subject to

0.9x1 ⫹ 0.5x2 ⫽ 100

Using the Kuhn–Tucker conditions and the second-order optimality conditions, what can you conclude about the following points; (a) x1 ⫽ 101.94, x2 ⫽ 16.508 (b) x1 ⫽ 84.1, x2 ⫽ 48.62 (c) x1 ⫽ 111.111, x2 ⫽ 0 (d) x1 ⫽ 0, x2 ⫽ 200 Note: ƒ is not a convex function. 5.19. Consider the problem Maximize

ƒ(x) ⫽ 3x1 ⫹ 5x2 ⫺ 0.2x13 / 1 ⫺ 0.05x22

Subject to

0.9x1 ⫹ 0.5x2 ⭐ 100 x1, x2 ⬎ 0

Solve using Everett’s generalized lagrange multiplier method. You may terminate after four iterations. Warning: This is a maximization problem. Be careful in updating Everett’s multiplies. 5.20. State whether each of the following functions is convex, pseudoconvex, quasi-convex, or strictly quasi-convex.

PROBLEMS

(a) ƒ1(x) ⫽ (b) (c) (d) (e)

再

0 ⫺(x ⫺ 1)2

259

冎

for 0 ⭐ x ⭐ 1 for 1 ⭐ x ⭐ 2

ƒ2(x) ⫽ ⫺x2 for x ⭓ 0 ƒ3(x) ⫽ ⫺x2 ⫺ x x⭓0 ƒ4(x) ⫽ x2 for 0 ⭐ x ⭐ 1 ƒ5(x) ⫽ ⫺x for 0 ⭐ x ⭐ 1

5.21. Consider the following functions: ƒ1(x) ⫽ ⫺x

ƒ2(x) ⫽ x ⫹ x3

Show that both ƒ1 and ƒ2 are pseudoconvex functions but the sum ƒ1 ⫹ ƒ2 is not a pseudoconvex function.

6 TRANSFORMATION METHODS In this chapter we begin our discussion of practical methods of solution for the NLP: Minimize

ƒ(x)

x 僆 RN

Subject to gj (x) ⭓ 0 hk (x) ⫽ 0

(6.1)

j ⫽ 1, 2, 3, . . . , J

(6.2)

k ⫽ 1, 2, 3, . . . , K

(6.3)

and (u) x (l) i ⭐ xi ⭐ x i

i ⫽ 1, 2, 3, . . . , N

(6.4)

We shall assume throughout that an initial estimate x (0) of the solution x* is available. This initial estimate may or may not be feasible, that is, satisfy Eqs. (6.2)–(6.4). We discuss algorithms that generate a sequence of points in RN from x (0) to x (T), where x (t) is the generic point and x (T), the limit point, is the best estimate of x* produced by the algorithm. The points x (t), t ⫽ 1, 2, 3, . . . , T, approximate stationary points of an associated unconstrained function called a penalty function. The original constrained problem is transformed into a sequence of unconstrained problems via the penalty function. The structure of the penalty function along with the rules for updating the penalty parameters at the close of each unconstrained minimization stage define the particular method. The penalty function is exact if only one unconstrained minimization is required. The Lagrange and Kuhn–Tucker theory of the previous chapter provides a powerful motivation and theoretical basis for the methods considered here. We have seen in the previous Lagrange discussion that it is useful to form an 260


6.1

PENALTY CONCEPT

261

unconstrained (Lagrangian) function in order to develop conditions of optimality. In addition, Everett [1] has given a method involving alternating Lagrange multiplier estimation and unconstrained minimization stages. The primary difficulty with the Everett approach is the numerical difficulty associated with the Lagrange multiplier estimation. On the other hand, we are obviously motivated to use these same (or similar) unconstrained functions to develop recursions, which are used iteratively to locate the constrained minimum. The idea of converting the constrained problem into a sequence of appropriately formed unconstrained problems is very appealing, since unconstrained problems can be solved both efficiently and reliably. We hope, of course, that only a few unconstrained subproblems of moderate difficulty will be required to approximate the constrained solution with acceptable accuracy. Penalty function methods are classified according to the procedure employed for handling inequality constraints, since essentially all transformation methods treat equality constraints in the same way. We will refer to these methods as interior or exterior point methods depending upon whether the sequence x (t) contains feasible or infeasible points, respectively. When the sequence of stationary points contains both feasible and infeasible points, we say the method is mixed.

6.1

PENALTY CONCEPT

Consider the penalty function P(x, R) ⫽ ƒ(x) ⫹ ⍀(R, g(x), h(x))

(6.5)

where R is a set of penalty parameters and ⍀, the penalty term, is a function of R and the constraint functions. The exact way in which the penalty parameters and the constraint functions are combined and the rules for updating the penalty parameters specify the particular method. As we have said, these methods are classified according to the manner in which they handle inequality constraints. In every case ⍀ is constructed so that solution of the sequence of associated subproblems produces a good estimate of the NLP. An interior point method is the result of selecting a form for ⍀ that will force stationary points of P(x, R) to be feasible. Such methods are also called barrier methods, since the penalty term forms a barrier of infinite P function values along the boundary of the feasible region. Any useful transformation technique should have the following desirable characteristics: 1. The subproblem solutions should approach a solution of the NLP, that is, limt→T⬍⬁ x (t) ⫽ x*. 2. The problem of minimizing P(x, R) should be similar in difficulty to minimizing ƒ(x). That is, the method will be less than useful if the

262

TRANSFORMATION METHODS

unconstrained subproblems are excessively difficult to solve, no matter how strong the theoretical basis of convergence. 3. R(t⫹1) ⫽ F(R(t)) should be simple. It seems reasonable to hope that the calculation overhead associated with updating the penalty parameters should be small compared to the effort associated with solving the unconstrained subproblems. (Note: This may in fact not be desirable for problems with very complex objective and constraint functions. In this case, considerable effort in updating the penalty parameters may be justified.) 6.1.1

Various Penalty Terms

In this section we consider a number of penalty forms that have been used widely and represent different procedures for handling constraints in an unconstrained setting. First consider the parabolic penalty used for equality constraints shown in Figure 6.1: ⍀ ⫽ R{h(x)}2

(6.6)

Notice that the parabolic penalty term equally discourages positive or negative violations of h(x). In addition, it is clear that with increasing values of R the stationary values of P(x, R) will approach x*, since in the limit as R grows large, h(x (T)) ⫽ 0. Note also that ⍀ is continuous and has continuous derivatives. Next consider several penalty forms for inequality constraints. Possibly the simplest form is the infinite barrier shown in Figure 6.2. This term assigns an infinite penalty to all infeasible points and no penalty to feasible points. The use of this term causes P(x, R) to be discontinuous along the boundary of the feasible region, and, rather obviously, ⵜP does not exist along the

Figure 6.1. Parabolic penalty.

6.1

PENALTY CONCEPT

263

Figure 6.2. Infinite-barrier penalty.

boundary. Infinite penalties are not machine realizable, but one can use a large positive constant consistent with the computer word length. Siddall [2, 3] and other investigators [4] have used this penalty form with some degree of success; for instance, Siddall suggests ⍀ ⫽ 1020

冘兩g (x)兩

(6.7)

j

j僆J

where in this case J identifies the set of violated constraints, that is, gj (x) ⬍ 0

for all

j僆J

(6.8)

Another useful form is the log penalty, as shown in Figure 6.3: ⍀ ⫽ ⫺R ln[g(x)]

(6.9)

Note that the term produces a positive penalty for all x such that 0 ⬍ g(x) ⬍ 1 and a negative penalty for all x such that g(x) ⬎ 1. In this way, interior points are artificially favored over boundary points. The negative penalties could be avoided by setting ⍀ ⫽ 0 for all x such that g(x) ⬎ 1, but in so

Figure 6.3. Log penalty.

264


doing we introduce a discontinuity in ⵜP in the interior region near the boundary. The log penalty is a barrier function and in fact is not defined for x such that g(x) ⬍ 0. A special recovery procedure is needed to handle infeasible points such as, say x (0). Contrary to various reports in the literature, it is possible to generate an infeasible point using such a barrier function, for instance with an overly ambitious first step in a line search, and therefore special logic must be included to detect, prevent, and/or recover from such a possibility. Beginning with an initial feasible point and a positive value for R (say, 10 or 100), R is decreased after each unconstrained minimization and approaches zero in the limit. The negative penalty is not present with the inverse-penalty term shown in Figure 6.4:

冋册

⍀⫽R

1 g(x)

(6.10)

The inverse-penalty term is also a barrier form, with similar difficulties associated with the evaluation of infeasible points. Rather obviously, a negative penalty is assigned to infeasible points, and special safeguards are in order. On the other hand, feasible points near the boundary are assigned quickly decreasing penalties as the interior region is penetrated. The terms P(x, R) and ⵜP do not exist along the boundary of the feasible region. Beginning with an initial feasible point and R positive, R is decreased toward zero in the limit. Finally consider the bracket operator term (Figure 6.5), ⍀ ⫽ R具g(x)典2

where

Figure 6.4. Inverse penalty.

(6.11)

6.1

PENALTY CONCEPT

265

Figure 6.5. Bracket operator.

具␣典 ⫽

再

␣ 0

if ␣ ⭐ 0 if ␣ ⬎ 0

(6.12)

Note first that the bracket operator term produces an exterior penalty function; stationary points of P(x, R) may indeed be infeasible. On the other hand, note that feasible and infeasible points are handled equally well by the term, and in fact no penalty is assigned to boundary or feasible points. The bracket operator produces an appealing penalty form. Here P(x, R) exists everywhere (for finite x) and is continuous. The value of R is chosen positive and is increased after each unconstrained stage. Let us demonstrate the use of these penalty terms with the following simple examples. Example 6.1

Parabolic Penalty Form

Minimize

ƒ(x) ⫽ (x1 ⫺ 4)2 ⫹ (x2 ⫺ 4)2

Subject to h(x) ⫽ x1 ⫹ x2 ⫺ 5 ⫽ 0 First form the penalty function using the quadratic penalty term: P(x, R) ⫽ (x1 ⫺ 4)2 ⫹ (x2 ⫺ 4)2 ⫹ R(x1 ⫹ x2 ⫺ 5)2 Now investigate the stationary points of P(x, R) as a function of R.

(6.13)

266


⭸P ⫽ 2(x1 ⫺ 4) ⫹ 2R(x1 ⫹ x2 ⫺ 5) ⫽ 0 ⭸x1 ⭸P ⫽ 2(x2 ⫺ 4) ⫹ 2R(x1 ⫹ x2 ⫺ 5) ⫽ 0 ⭸x2

x1 ⫽ x2 ⫽

5R ⫹ 4 2R ⫹ 1

(6.14)

Is the method convergent? L’Hospital’s rule produces the following limit: lim R→⬁

5R ⫹ 4 5 ⫽ ⫽ 2.5 2R ⫹ 1 2

(6.15)

The answer is yes, since by inspection of Figure 6.6 we see that x* ⫽ [2.5, 2.5]T, with ƒ(x) ⫽ 4.5. The stationary values of P(x, R) for various values of R are given in Table 6.1. Figures 6.7–6.9 give contours of P(x, R) for various values of R. Notice that the contours of P(x, R) do indeed change to align with h(x) as R grows large. This distortion of the shape of the contours of P(x, R) has a positive effect, in that it is directly responsible for the convergence of the method. Let us pause to examine the stationary point of P(x, 1) in Example 6.1. Obviously it is not x* but is reasonably close. Note that x (t) ⫽ [3, 3,]T;

Figure 6.6. Two-variable equality-constrained problem.

6.1

Table 6.1

PENALTY CONCEPT

267

Stationary Values of Parabolic Penalty

R

(t) x(t) 1 ⫽ x2

ƒ(x(t))

h(x(t))

P(x(t), R)

0 0.1 1 10 100

4.0000 3.7500 3.0000 2.5714 2.5075 2.5000

0.0000 0.1250 2.0000 4.0818 4.4551 4.5000

3.0000 2.5000 1.0000 0.1428 0.0150 0.0000

0.0000 0.7500 3.0000 4.2857 4.4776 4.5000

⬁

ƒ(x (t)) ⫽ 2.0 and h(x (t)) ⫽ 1.0. In some engineering applications this answer might be close enough, but in general we would be concerned about the magnitude of the violation of h(x). The question then arises, how shall we change the penalty function such that its stationary point is closer to x*? Note that in finding the stationary point of P(x, 1) exactly we have achieved a point that has too low a value for ƒ(x) and a corresponding violation of h(x). But we built this difficulty into P(x, 1) when we chose the penalty parameter R ⫽ 1, since R tends to weight the equality constraint against the objective. That is, as R grows large, the stationary point of P(x, R) will tend to satisfy h(x) more closely and approach x*, as can be seen in Table 6.1. Obviously,

Figure 6.7. Parabolic penalty contours, R ⫽ 1.

268



if we choose R ⫽ ⬁ we produce the exact solution in one unconstrained minimization. Hold it! How will we choose R ⫽ ⬁? Of course we cannot, but we could choose R to be very large. The next problem then is to decide how large. In Table 6.1 we see that R ⫽ 100 produces the correct solution to two places when exact minimization is employed. But exact minimization will seldom be possible, and one of the unconstrained search methods previously considered will be employed. We are tempted to choose an even larger value for R so that the error in the result due to R and the error due to inexact minimization will tend to add to an acceptable value. But there is a very real practical difficulty with this approach, associated with the relationship R has with the shape of the contours of P(x, R). We see as R grows large, the contours of P become more and more distorted. Therefore the unconstrained minimization may become increasingly difficult. This thinking suggests a multistage strategy for approximating the solution. First choose R to be a relatively small value, say 0.1, and minimize P(x, 0.1); then increase R to 1.0 and once again minimize P(x, 1.0), using the result from P(x, 0.1) to begin. We continue this increase in R and solution of the resulting subproblem until we notice acceptable small changes in the sequences x (t), ƒx (t), and P(x (t), R).

6.1

PENALTY CONCEPT

269


We hope that the estimate of the minimizing point from the previous stage will be close enough to the minimizing point for the new stage to offset the increased minimization difficulty due to the increase in R. This will often be the case if we do not increase R too rapidly, but when R is increased slowly, many stages are required and the overall computing time to the solution can be great. This is the major difficulty with all penalty methods, and much effort has been expended to develop rules for updating R to increase the overall efficiency of the approach. We shall see that very similar difficulties occur when inequalities are present. It is of interest to examine the performance of several of the inequality constraint penalty forms on a similar problem. Example 6.2

Inequality Penalty Forms

Let us change the constraint in Example 6.1 to an inequality form such that x* remains the same: g(x) ⫽ 5 ⫺ x1 ⫺ x2 ⭓ 0

(6.16)

We handle this problem in a way that is very similar to the equality constraint case by employing the bracket operator:

270


P(x, R) ⫽ ƒ(x) ⫹ R具g(x)典2

(6.17)

P(x, R) ⫽ (x1 ⫺ 4)2 ⫹ (x2 ⫺ 4)2 ⫹ R具5 ⫺ x1 ⫺ x2典2

(6.18)

or

Next we investigate the stationary points of P(x, R): ⭸P ⫽ 2(x1 ⫺ 4) ⫹ 2R具5 ⫺ x1 ⫺ x2典(⫺1) ⫽ 0 ⭸x1 ⭸P ⫽ 2(x2 ⫺ 4) ⫹ 2R具5 ⫺ x1 ⫺ x2典(⫺1) ⫽ 0 ⭸x2

which leads to x1 ⫽ x2

(6.19)

(x1 ⫺ 4) ⫺ R具5 ⫺ 2x1典 ⫽ 0

(6.20)

as before. Therefore,

Assume that the argument of the bracket operator is equal to, greater than, and less than zero, respectively, and analyze the results. Therefore assume 2x1 ⭓ 5, which implies (x1 ⫺ 4) ⫺ R(5 ⫺ 2x1) ⫽ 0 or x1 ⫽

5R ⫹ 4 ⫽ x2 2R ⫹ 1

(6.21)

Therefore limR→⬁ x1 ⫽ 2.5, which agrees with the assumption. Therefore for P(x, 1) the stationary point is x (t) ⫽ [3, 3]T, with ƒ(x (t)) ⫽ 2, g(x (t)) ⫽ ⫺1, and P(x, 1) ⫽ 3. As R goes from 0 to ⬁, the stationary points of P(x, R) travel along the straight line from [4, 4]T, the unconstrained solution, to the constrained solution at [2.5, 2.5]T. For all values of R the stationary points are infeasible, and hence this is an exterior form. Comparing Eqs. (6.21) and (6.15) we see that the path to the solution with exact minimization is exactly the same for the parabolic and bracket operator forms. Plots of the contours of P(x, R) for R ⫽ 1, 10, 100 are given in Figures 6.10–6.12. Consider now the application of the log penalty to this same inequality problem:

6.1

PENALTY CONCEPT

Figure 6.10. Bracket operator penalty contours, R ⫽ 1.


271

272



P(x, R) ⫽ ƒ(x) ⫺ R ln(g(x))

(6.22)

P(x, R) ⫽ (x1 ⫺ 4)2 ⫺ R ln(5 ⫺ x1 ⫺ x2)

(6.23)

or

The stationary conditions are

冋冋

⫽0

⭸P 1 ⫽ 2(x2 ⫺ 4) ⫹ R ⭸x2 5 ⫺ x2 ⫺ x2

⫽0

which leads to x1 ⫽ x2 Therefore,

册册

⭸P 1 ⫽ 2(x1 ⫺ 4) ⫹ R ⭸x1 5 ⫺ x1 ⫺ x2

6.1

2(x1 ⫺ 4) ⫹

冋

PENALTY CONCEPT

册

R ⫽0 5 ⫺ 2x1

273

(6.24)

This leads to 2x21 ⫺ 13x1 ⫹ 20 ⫺

R ⫽0 2

(6.25)

and the roots produce the desired stationary values: 13 1 x1 ⫽ –– 4 ⫺ – 4 兹9 ⫹ 4R

(6.26)

where we reject the other root since it is infeasible. Now assume that R begins large and is incremented toward zero: lin x1 ⫽ 2.5

ƒ(x (T)) ⫽ 4.5

R→0

Once again we see that the method converges with exact minimization. The stationary values of P(x, R) for various values of R are given in Table 6.2, and plots of the contours of P(x, R) are given in Figures 6.13–6.15. Notice from Table 6.2 that P(x (t), R) does not increase monotonically to ƒ(x*) but increases and then decreases. This is due to the negative-penalty effect included in the log term. Finally we consider the inverse penalty:

冋册

P(x, R) ⫽ ƒ(x) ⫹ R

1 g(x)

(6.27)

冋

P(x, R) ⫽ (x1 ⫺ 4)2 ⫹ (x2 ⫺ 4)2 ⫹ R

1 5 ⫺ x1 ⫺ x2

册

(6.28)

As before, examine the stationary values of P(x, R) for various values of R:

Table 6.2

Stationary Values of Log Penalty

R

x1 ⫽ x2

ƒ(x)

g(x)

⫺R ln(g(x))

P(x, R)

100 10 1 0.1 0.01 0

⫺1.8059

67.4170 12.5000 5.4542 4.5995 4.5100 4.5000

8.6118 2.0000 0.3028 0.0034 0.0034 0.0000

⫺215.3133 ⫺6.9315 ⫹1.1947 ⫹0.3411 ⫹0.0568

⫺147.8963

1.5000 2.3486 2.4835 2.4983 2.5000

0.0000

5.5685 6.6489 4.9406 4.5668 4.5000

274


Figure 6.13. Log penalty contours, R ⫽ 100.

⭸P ⫽ 2(x1 ⫺ 4) ⫹ ⭸x1 ⭸P ⫽ 2(x2 ⫺ 4) ⫹ ⭸x2

冋冋

册册

R ⫽0 (5 ⫺ x1 ⫺ x2)2 R ⫽0 (5 ⫺ x1 ⫺ x2)2

which leads to x1 ⫽ x2 and 4x31 ⫺ 36x21 ⫹ 105x1 ⫺ 100 ⫹

R ⫽0 2

(6.29)

The stationary points are defined by the zeros of this nonlinear function. We use a zero-finding algorithm [5] to find the values given in Table 6.3, and plots of the contours of P(x R) for various R values are in Figures 6.16–6.18. Once again we see that the stationary values of the penalty function approach x* as R → 0 with exact minimization.

6.1

PENALTY CONCEPT



275

276


Table 6.3

Stationary Values of Inverse Penalty

R

x1 ⫽ x2

ƒ(x)

g(x)

R / g(x)

P(x, R)

100 10 1 0.1 0.01 0.001 0

0.5864 1.7540 2.2340 2.4113 2.4714 2.4909 2.5000

23.3053 10.0890 6.2375 5.0479 4.6732 4.5548 4.5000

3.8272 1.4920 0.5320 0.1774 0.0572 0.0182 0.0000

26.1288 6.7024 1.8797 0.5637 0.1748 0.0549 0.0000

49.4341 16.7914 8.1172 5.6116 4.8480 4.6097 4.500

Consider the contours of the inverse-penalty function shown in Figures 6.16–1.18. We see in Table 6.3 that the method converges along an interior path as R is decreased. Now look carefully at Figures 6.17 and 6.18. In Figure 6.18 we see the appearance of a false minimum of the inverse-penalty function. Imagine the result if our starting point is (5,5). Notice that we would

Figure 6.16. Inverse-penalty contours, R ⫽ 100.

6.1

PENALTY CONCEPT

277


need additional logic to recover from the false minimum once it is encountered. This should be clear from examination of Figure 6.18 and the logic of the method. The examples demonstrate both the utility of the various standard penalty approaches and their respective weaknesses. That is, we have seen that convergence is associated with ever-increasing distortion of the penalty contours, which increases significantly the possibility of failure of the unconstrained search method. Recall that we have assumed in our discussion of convergence that the unconstrained searches could be completed successfully. Any inherent tendency that opposes this assumption seriously undermines the utility of the penalty function approach. 6.1.2

Choice of Penalty Parameter R

With each of the parametric penalty forms considered, it is necessary to choose an initial value for R and to adjust this value after each unconstrained search to force convergence of the sequence of stationary points x (t). Hope-

278



fully it is clear that R must be chosen so as to increase the weight of constraint violations for exterior forms and to decrease the weight for interior or barrier forms from stage to stage. In addition, we must devise a strategy for incrementing R from stage to stage. For instance, with the simple parabolic penalty for equality constraints it would seem reasonable to begin with R ⫽ 0, that is, conduct an unconstrained search of ƒ(x) and increase R by ⌬R (say, powers of 10) at each stage. This would force the stationary points to approach satisfaction of the equality constraint. Experimentally we observe that the stationary points x (t)(R) converge smoothly to x* (with exact searches). Fiacco and McCormick [6, 7] and Lootsma [8, 9] have suggested that an approximating function be fitted to the sequence x (t)(R) and that this approximation be evaluated at R ⫽ 0 to provide a better estimation of x (t⫹1) to begin the next unconstrained search. It is hoped that in this way ⌬R can be larger than would otherwise be possible. Fletcher and McCann [10] have conducted tests that suggest that a quadratic approximating function is best. Even though this extrapolation technique may improve the performance of the basic penalty approach, it cannot remove the distortion in the shape of the contours of P(x, R) as R is adjusted. In fact, it is this very distortion that

6.2

ALGORITHMS, CODES, AND OTHER CONTRIBUTIONS

279

produces convergence of the method (with exact solution to the subproblems). On the other hand, the change in the penalty parameters forces the subproblems to be progressively less well-conditioned. That is, for every penalty form considered to this point the unconstrained subproblems will automatically become ill-conditioned as a natural consequence of the changes in R needed to force convergence. This fact is easily observed from the shape of the contours in the previous figures and has been noted by several investigators, including Murray [11] and Lootsma [12]. In fact, Lootsma shows that M (the number of active constraints at x*) eigenvalues of ⵜ2P(x, R) vary with 1/R, which implies that ⵜ2P(x, R) becomes increasingly ill-conditioned as R approaches zero. This means that the unconstrained subproblems become increasingly difficult to solve, and termination may be a result of failure of the unconstrained method rather than convergence of the penalty method. Once again this is purely a numerical problem, since very strong convergence results can be demonstrated in the presence of exact calculations.

6.2


It is really straightforward to build a useful algorithm with the concepts previously discussed. A simple penalty function algorithm would take the following form. Penalty Function Algorithm Step 1. Define N, J, K, ␧1, ␧2, ␧3, x (0), and R(0), where ␧1 ⫽ line search termination criterion ␧2 ⫽ unconstrained method termination criterion ␧3 ⫽ penalty termination criterion x (0) ⫽ initial estimate of x* R(0) ⫽ initial set of penalty parameters Step 2. Form P(x, R) ⫽ ƒ(x) ⫹ ⍀(R, g(x), h(x)). Step 3. Find x (t⫹1) such that P(x (t⫹1), R(t)) → min, with R(t) fixed. Terminate, using ␧2, and use x (t) to begin next search. Step 4. Is 兩P(x (t⫹1), R(t)) ⫺ P(x (t), R(t⫺1))兩 ⭐ ␧3? Yes: Set x (t⫹1) ⫽ x (T) and terminate No: Continue. Step 5. Choose R(t⫹1) ⫽ R(t) ⫹ ⌬R(t) according to a prescribed update rule, and go to 2. Certainly a useful code would include additional steps and logic. For instance, it would seem wise to monitor several sequences x (t), R(t), g(x (t)), h(x (t)),

280


ƒ(x (t)) in addition to the one given in step 4 in order to make a termination decision. Possibly the first record of a penalty-type method in the literature is the work of Courant [13], but it is the ‘‘created response surface’’ idea of Carroll [14] that marks the beginning of the modern era of transformation techniques. The work of Fiacco and McCormick [15–17] is also important because it provided the first widely available NLP code, SUMT (which we discuss further in the next section). Somewhat later (1967), Zangwill [18] considered the general utility of the penalty approach and provided conditions of convergence for various penalty forms. One of the most complete examinations of penalty methods has most probably been provided by Lootsma [19–21], who in particular has demonstrated the advantages and disadvantages of the various forms by examination of their properties along the boundary of the feasible region. In addition, McCormick [22] gives an interesting discussion of the relative merits of penalty and nonpenalty methods for constrained problems. For reasons previously discussed at length, we might expect to encounter difficulties in using some or all of the line search techniques from Chapter 2 to minimize P(x, R) along a prescribed direction, since ⵜP(x, R) will take on extremely high values near the boundary. Many have recognized this difficulty, and Lasdon [23] has proposed an efficient technique especially designed to minimize penalty-type functions. It is not possible in this brief discussion to give a proper survey of the rich history of penalty methods; instead we have mentioned some of the more notable contributions. Much more complete surveys are given by Lootsma [4], Moe [24], Ryan [25], Fox [26], Himmelblau [27], and Avriel [28]. A number of penalty-type codes have been widely used; these include SUMT and OPTISEP, which are discussed shortly; COMPUTE II [29], a Gulf Oil Corporation software package; a package developed by Afimiwala [30] at the State University of New York at Buffalo; and a package developed by Lill [31]. In every case a variety of penalty forms and unconstrained searching methods are available to the user. The utility of a number of the standard penalty forms is described in a comparative study by DeSilva and Grant [32], and the value of the approach is demonstrated by various applications given by Bracken and McCormick [33]. Rao and Gupta [34, 35] have employed penalty methods in the design of thermal systems. Possibly the two most widely used penalty function packages are SUMT [36] by Fiacco, McCormick, and co-workers and OPTISEP [3] by Siddall. Sandgren [37] gives availability information and test results for these codes. The SUMT package has been released in at least two forms:

冘冋g 1(x)册 ⫹ R1 冘 [h (x)] 1 P(x, R) ⫽ ƒ(x) ⫺ R 冘 ln[g (x)] ⫹ 冘 [h (x)] R J

1967:

K

P(x, R) ⫽ ƒ(x) ⫹ R

2

k

j⫽1 J

1970:

K

2

j

j⫽1

(6.30)

k⫽1

j

k

k⫽1

(6.31)

6.2


281

where R(0) ⬎ 0 and R(0) ⬎ R(1) ⬎ R(2) → 0. The 1970 version (SUMT-IV) is the one Sandgren tested (see Chapter 12). This package contains four separate penalty function algorithms, all employing the penalty form in (6.31) but with different rules for updating R and/or performing the unconstrained search. In particular, the user can choose the Cauchy, DFP, or one of two forms of the modified Newton method for the unconstrained searches. The penalty parameter R is decreased after each unconstrained subproblem, and the extrapolation technique previously discussed is employed as well. The reader is referred to the SUMT-IV user’s manual for more specific details. The OPTISEP package contains a number of penalty methods, which differ primarily in the unconstrained method used. Two penalty forms are used:

冘兩具g (x)典兩 ⫹ 10 冘兩h (x)兩 J

P(x, R) ⫽ ƒ(x) ⫹ 1020

K

20

j

j⫽1

k

(6.32)

k⫽1

and the form given in (6.30). SEEK1 employs (6.32) and a Hooke–Jeeves direct-search method. After termination, a number of trial points are generated randomly about x (T), and the search begins anew if a lower value of P(x, R) is found. SEEK3 uses the 1967 SUMT form in (6.30) and uses (6.31) to locate an initial feasible point. Hooke–Jeeves is again used, and R is reduced by a constant factor at each stage. SIMPLX employs (6.30) and the simplex direct-search method. DAVID uses (6.30) and the DFP gradient-based searching method. Finally, MEMGRD employs (6.30) and Miele’s memory gradient method, which is generalization of the Fletcher–Reeves conjugate gradient method. All programs use a similar format for problem entry. The transformation methods described in this section continue to be used in spite of their various weaknesses. Fletcher [38] discusses penalty and barrier methods in the hope of identifying their defects and making users aware of potential remedies. This is a particularly useful line of thinking, since it is highly probable that these methods will continue to be used, primarily because they are easy to understand and implement. Nocedal and Wright [39] examine modern nonlinear optimization methods including penalty methods. Franz et al. [40] has reported on extensive numerical results with these and related methods. He also provides a number of software packages [41]. The Optimization Technology Center [42], a cooperative venture between Argonne National Laboratory and Northwestern University, provides a number of very useful reports and links to a number of widely used software packages. OPTLIB [43], a software package written by G. A. Gabriele and designed to be used in an educational environment to help both instructors and students of modern optimization, contains all of the methods discussed in this chapter. Bertsekas [44] gives a particularly useful review of penalty methods, including exact penalty methods and augmented Lagrangian methods similar to those discussed in the next section.

282

6.3


METHOD OF MULTIPLIERS

It has been recognized for some time that the progressive ill-condition of the subproblems generated in the standard (SUMT-like) penalty approach limits the utility of the method for practical applications, particularly in competition with the linearization-based methods (to be discussed in Chapters 8 and 9). With this in mind, several investigators have suggested fixed-parameter SUMT-like penalty methods (most notably Kowalik [45], Fiacco and McCormick [46], Fiacco [47], Rosenbrock [48], Huard [49], and Staha [50] and other extensions in order to mitigate this malady to some extent. Of particular interest is Huard’s method of centers, where we maximize

写 g (x) J

P(x, t) ⫽ [t ⫺ ƒ(x)]

(6.33)

j

j⫽1

where t is a moving truncation at each maximization stage, say, t(t) ⫽ ƒ(x (t⫺1)), where x (t⫺1) is the maximum point of the previous stage. Note that t is adjusted in the normal course of the calculations. It has been shown [28] that Huard’s product form is equivalent to the parameterless penalty form:

冘 ln g (x) J

P(x) ⫽ ⫺ln[ƒ(x (t⫺1)) ⫺ ƒ(x)] ⫺

j

(6.34)

j⫽1

where minimization is implied. Avriel comments that these parameter-free methods, although attractive on the surface, are exactly equivalent to SUMT with a particular choice of updating rule for R. In fact, Lootsma [51] shows that the convergence rates for SUMT and the method of centers are quite comparable. Recall the role of the Lagrange function, L(x, u, v) ⫽ ƒ(x) ⫺

冘 u g (x) ⫺ 冘 v h (x) j j

j

k k

(6.35)

k

in the discussion of the Kuhn–Tucker optimality conditions of Chapter 5. We are tempted to seek the values of x, u, and v that cause the Lagrangian to take on a minimum value, but we do not, since we know that the solution (x*, u*, v*) is a stationary point of L but not a minimum (in fact, it is a saddlepoint of L). On the other hand, it is productive to augment the Lagrangian (by, e.g., a quadratic loss penalty term) to form an unconstrained function whose minimum is a Kuhn–Tucker point of the original problem. In this section we discuss such an augmented Lagrangian method, called the method of multipliers, which can also be viewed as a remedy to the contour distortion malady previously discussed.

6.3

6.3.1


283

Penalty Function

Consider the function

冘 {具g (x) ⫹ ␴ 典 ⫺ ␴ } J

P(x, ␴, ␶) ⫽ ƒ(x) ⫹ R

2

j

2 j

j

j⫽1

冘 {[h (x) ⫹ ␶ ] ⫺ ␶ } K

⫹R

2

k

k

2 k

(6.36)

k⫽1

where R is a constant scale factor (R may vary from constraint to constraint but remains constant from stage to stage), and the bracket operator is defined as before. The solution of the unconstrained subproblem is called a stage. The ␴j and ␶k parameters are fundamentally different from R in that they bias the penalty terms in a way that forces convergence of the iterates under rather mild conditions on the function. The method to be considered here is due to Schuldt [52, 53] and is similar to Rockafellar’s modification [54] of the Powell–Hestenes [55, 56] multiplier method for equalities only. We shall see that the method of multipliers (MOM) is quite similar in philosophy to Everett’s method [1], with the exception that the MOM adjusts the Lagrange multiplier estimates (␴ and ␶) quite conveniently and automatically. We drop the R in the penalty notation to underline the fact that R is constant from stage to stage; only the ␴ and ␶ elements change. Further more, ␴ and ␶ are constant vectors during each unconstrained minimization. The updating rules (to be explained next) are applied to the elements of ␴ and ␶ to prepare a new penalty function for the next unconstrained stage. It is not necessary for the starting vector x (0) to be feasible, and a convenient but completely arbitrary choice of parameters for the first stage is ␴ ⫽ ␶ ⫽ 0. Thus the first minimization stage is identical to the first unconstrained minimization using standard exterior penalty terms previously discussed. 6.3.2

Multiplier Update Rule

Suppose that the vector x (t) minimizes the tth-stage penalty function:

冘 {具g (x) ⫹ ␴ J

P(x, ␴ (t), ␶ (t)) ⫽ ƒ(x) ⫹ R

j

典 ⫺ [␴ j(t)]2}

(t) 2 j

j⫽1

冘 {[h (x) ⫹ ␶ K

⫹R

k

2 ] ⫺ [␶ (t) k ] }

(t) 2 k

(6.37)

k⫽1

Multiplier estimates for the (t ⫹ 1)st stage are formed according to the following rules:

284


␴ (t⫹1) ⫽ 具gj (x (t)) ⫹ ␴ (t) j j 典

j ⫽ 1, 2, 3, . . . , J

(6.38)

␶ (t⫹1) ⫽ hk (x (t)) ⫹ ␶ (t) k k

k ⫽ 1, 2, 3, . . . , K

(6.39)

Because of the bracket operator, ␴ has no positive elements, whereas the elements of ␶ can take either sign. These parameters serve as a bias in the arguments of the penalty terms, and the updating rules tend to change the bias in a way that increases the penalty on violated constraints in successive stages, thus forcing the stationary points x (t) toward feasibility. 6.3.3

Penalty Function Topology

The effect of changes in the multipliers ␴ and ␶ upon the shape of the contours of P(x, ␴, ␶) is not obvious from (6.37) but can be seen in an expression for ⵜ2P(x), the Hessian matrix of the penalty function. Using the definition of the bracket operator, we see the gradient to be

冘 {具g (x) ⫹ ␴ 典ⵜg (x)} J

ⵜP(x) ⫽ ⵜƒ(x) ⫹ 2 R

j

j

j

j⫽1

冘 {[h (x) ⫹ ␶ ] ⵜh (x)} K

⫹ 2R

k

k

(6.40)

k

k⫽1

An additional differentiation produces the matrix of second derivatives:

冘 {具g (x) ⫹ ␴ 典 ⵜ g (x) ⫹ [ⵜg (x)] } J

ⵜ2P(x) ⫽ ⵜ2ƒ(x) ⫹ 2R

2

j

j

2

j

j

j⫽1

冘 {[h (x) ⫹ ␶ ] ⵜ h (x) ⫹ [ⵜh (x)] } K

⫹ 2R

2

k

k

2

k

(6.41)

k

k⫽1

It can be shown that the radius of curvature of a particular contour of P(x) varies as ⵜ2P(x). Let the constraints g(x) and h(x) be linear; then,

冘 {ⵜg (x)} ⫹ 2R 冘 {ⵜh (x)} J

ⵜ2P(x) ⫽ ⵜ2ƒ(x) ⫹ 2R

K

2

2

j

j⫽1

k

(6.42)

k⫽1

So we see that ⵜ2P(x) is independent of ␴ and ␶ when the constraints are linear, and therefore the contours do not change in shape from stage to stage in this case. The contours of P(x) are simply shifted or biased relative to the contours of ƒ(x). This is an obviously desirable property, since convergence of the method does not depend on distortion of the contours of P(x) and the associated infinite limit point of penalty parameters. Furthermore, we see that the contours of P(x) do change in shape when the constraints are nonlinear, but the effect is second order rather than primary as in the previously considered methods.

6.3

6.3.4


285

Termination of the Method

Progress of the method is checked by monitoring the sequences x (t), ␴ (t), ␶ (t), ƒ(x (t)), g(x (t)), and h(x (t)). It is assumed that the unconstrained algorithm can find the stationary points and has its own suitable criteria for termination (ending the stage). Termination occurs when the multiplier update and associated unconstrained minimization fail to produce change in one or more of the monitored sequences. Termination then produces the limit point x (T), which we now show to be a Kuhn–Tucker point. Note that the gradient of the penalty function must vanish at the end of a stage. Also, according to expressions (6.38) and (6.39), the quantities 具gj (x (t)) (t) (t) (t⫹1) ⫹ ␴ (t) and ␶ (t⫹1). Consider j 典 and [hk (x ) ⫹ ␶ k ] become, respectively, ␴ (T) (T) (T) (T) now the supposed limits ␴ , gj (x ), ␶ , and hk (x ). From (6.39) it is clear that ␶ (T) can exist only if hk (x (T)) ⫽ 0. Similarly, (6.38) implies that the limits ␴ (T) and gj (x (T)) exist only if gj (x (T)) ⬎ 0

and

␴ (T) ⫽0 j

(6.43)

gj (x (T)) ⫽ 0

and

␴ (T) ⭐0 j

(6.44)

or

Accordingly, the gradient given in (6.40) but evaluated at the limit point x (T) can be written equivalently:

冘␴ J

ⵜP(x (T)) ⫽ ⵜƒ(x (T)) ⫹ 2R

j⫽1

冘␶ K

(T) j

ⵜgj (x (T)) ⫹ 2R

(T) k

ⵜhk (x (T)) ⬅ 0

k⫽1

(6.45) with gj (x (T)) ⭓ 0

j ⫽ 1, 2, 3, . . . , J

(6.46)

(T) ␴ (T) )⫽0 j gj (x

j ⫽ 1, 2, 3, . . . , J

(6.47)

␴ j(T) ⭐ 0

j ⫽ 1, 2, 3, . . . , J

(6.48)

hk (x (T)) ⫽ 0

k ⫽ 1, 2, 3, . . . , K

(6.49)

Notice that expressions (6.45)–(6.49) are a restatement of the Kuhn–Tucker conditions gives in Chapter 5, and therefore the limit point x (T) is a Kuhn– Tucker point. Furthermore, we see from (6.45) that

286


uj ⫽ ⫺2R␴ (t) j

(6.50)

vk ⫽ ⫺2R␶ k(t)

(6.51)

are Lagrange multiplier estimates and are obtained with essentially no additional calculation. The dual-variable nature of these Lagrange multiplier estimates has been studied by a number of investigators, including Gould [57] and Gill [58]. 6.3.5

MOM Characteristics

In summary, we note the following general characteristics of the method of multipliers: 1. limt→T⬍⬁ x (t) ⫽ x*, a Kuhn–Tucker point. We have shown that this result follows assuming only that the unconstrained minimizations terminate. We expect the unconstrained subproblems to be less ill-conditioned than with the previously considered methods, because of the structure of the penalty form and the nature of the ␴ and ␶ parameters. This leads one to observe that 2. Minimize P(x) is similar in difficulty to Minimize ƒ(x) for reasonable values of R. 3. Finally we note that the ␴ and ␶ update rules are simple, requiring essentially no additional calculation. Root [59] and others [60, 61] have investigated the utility of more complex update rules for the multiplier estimates but have found in general that the additional computational overhead outweighs the decrease in the number of stages to termination. As previously noted, the value of this result is a strong function of the relative computational effort required to compute the various problem functions. In addition, Root [59] has given an ‘‘interiorlike’’ version of the method of multipliers and shown it to possess many of the positive characteristics of the exterior form considered here, with the additional pleasant feature that the iterates x (t) tend to remain in the feasible region. On the other hand, the formulation considered here does have its weaknesses. As mentioned, the sequence of iterates x (t) will almost certainly approach x* from the infeasible region. In addition, we have the problem of choosing R (which we discuss in Section 6.3.6), and recall our assumption of convergence of the unconstrained subproblems. Kort [61] shows the method to be globally convergent when applied to convex programs and gives a termination rule for the subproblems that preserves global convergence and rate of convergence with inexact subproblem solutions.

6.3

Example 6.3


287

Example 6.1 Revisited

Consider once again the equality-constrained problem given in Example 6.1: Minimize

ƒ(x) ⫽ (x1 ⫺ 4)2 ⫹ x2 ⫺ 4)2

Subject to h(x) ⫽ x1 ⫹ x2 ⫺ 5 ⫽ 0 P(x, ␶) ⫽ (x1 ⫺ 4)2 ⫹ (x2 ⫺ 4)2 ⫹

1 1 (x ⫹ x2 ⫺ 5 ⫹ ␶)2 ⫺ ␶ 2 R 1 R

⭸P 2 ⫽ 2(x1 ⫺ 4) ⫹ (x1 ⫹ x2 ⫺ 5 ⫹ ␶) ⫽ 0 ⭸x1 R ⭸P 2 ⫽ 2(x2 ⫺ 4) ⫹ (x1 ⫹ x2 ⫺ 5 ⫹ ␶) ⫽ 0 ⭸x2 R

These expressions produce x1 ⫽ x2 ⫽

5 ⫹ 4R ⫺ ␶ 2⫹R

(6.52)

For convenience, choose R ⫽ 1, to get x1 ⫽ x2 ⫽ 3 ⫺

␶ 3

(6.53)

Now we use (6.39) to update the multiplier ␶ at the end of each constrained minimization of P(x, ␶), which is here done exactly using (6.52). The results of these calculations are given in Table 6.4. Note that P(x, ␶) increases as the iterations proceed. We are, in fact, locating a saddlepoint of P(x, ␶), since we are minimizing with respect to x but maximizing with respect to ␶. Contour maps of P(x, ␶) for various values of ␶ are given in Figures 6.19–6.21. Once again, since the constraint is linear, we have no change in the shape of the contours of P(x, ␶) as ␶ converges to its finite terminal value. Table 6.4

MOM Results

t

␶

x1 ⫽ x2

ƒ(x)

P(x, ␶)

0 1 2 3 4 T

0 1.0000 1.3333 1.4444 1.4818 1.5000

3.0000 2.6667 2.5556 2.5185 2.5062 2.5000

2.0000 3.5556 4.1728 4.3896 4.4629 4.5000

3.0000 4.3333 4.4814 4.4979 4.4997 4.5000

288


Figure 6.19. MOM contours, ␶ ⫽ 1.0.

13 Figure 6.20. MOM contours, ␶ ⫽ –– 9.

6.3


289

Figure 6.21. MOM contours, ␶ ⫽ 1.5.

6.3.6

Choice of R-Problem Scale

Poor problem scale is possibly the most frequently encountered difficulty in the application of modern transformation methods in engineering practice. A problem is poorly scaled (or ill-scaled) if the moduli of the variables and the problem functions are not comparable. When this occurs, one or more of the design variables and/or constraints can artificially dominate the search and cause premature termination of any algorithm. In Chapter 13 we discuss a method by Root [62] that senses ill-scaling and automatically conditions the problem so that the minimization algorithm sees a better-conditioned problem. Keefer and Gottfried [63, 64] proposed a similar method in 1970 for use with the standard penalty approach, whereas Root was specifically interested in enhancing the performance of the MOM approach. Even though both methods were developed for use with transformation methods, they could be used to condition problems to be solved by any algorithm. It is worthy of note that the very nature of transformation methods makes problem scale an important issue. Accordingly, some effort to scale the problem variables and constraints seems advisable. 6.3.7

Variable Bounds

Most transformation approaches ignore the special nature of the variable bounds given in (6.4) and treat them as inequality constraints. Accordingly,

290


for each design variable we must add two inequality constraints. Admittedly the additional constraints are simple, but a more direct approach seems desirable. We seek a procedure that handles the bounds directly and does not increase the dimension (number of constraints) of the problem. The technique must be able to recognize that one or more variables is at or exceeds its bounds and hold these variables on bounds if violation would otherwise occur. In addition, the technique must conveniently sense and allow feasible changes in bound variables. One can imagine many approaches that would satisfy these general requirements. A particularly interesting questions is the action to be taken when one or more bounds are violated. Since the unconstrained searches are typically executed using a direction-generating method, we might simply back up along the current search direction until all variables are within bounds. This would preserve the line of search in the unconstrained algorithm but would require a zero-finding operation and would therefore be computationally expensive. A more direct, less elegant, and less expensive approach would be to simply set the out-of-bounds variables to their violated bounds simultaneously. Obviously this destroys the search direction, but our experience suggests that this is the best approach. A related issue is the possibility of exploiting the presence of linear constraints. When some of the constraints are linear, certain economies of calculation are forthcoming. First, the rate of convergence will be enhanced due to the favorable impact on the penalty function topology. Second, the line search algorithm can be made more efficient, since the linear boundaries are

Figure 6.22. Welded beam.

6.3


291

trivial to locate. In our own experiments with the MOM [62] we have included scaling and variable-bound algorithms but we have not explored the efficiencies associated with linear constraint exploitation. Example 6.4

Welded Beam

Consider once again the welded beam problem [65] described in Chapter 1 and shown in Figure 6.22. Here we examine the solution using a MOM implementation called BIAS [66]. BIAS uses the simple multiplier update rule, scaling algorithm, and variable-bound procedure previously discussed. The objective remains the same, but three of the constraints (g4, g5, and g7) are simply bounds on variables and are entered as such. The problem then becomes Minimize

ƒ(x) ⫽ (1 ⫹ c3)x12x2 ⫹ c4x3x4(L ⫹ x2)

(6.54)

Subject to g1(x) ⫽ ␶d ⫺ ␶ (x) ⭓ 0 g2(x) ⫽ ␴d ⫺ ␴ (x) ⭓ 0 g3(x) ⫽ x4 ⫺ x1 ⭓ 0

(6.55)

g4(x) ⫽ Pc(x) ⫺ F ⭓ 0 g5(x) ⫽ 0.25 ⫺ ⭸(x) ⭓ 0 0.125 ⭐ x1 ⭐ 10 0.100 ⭐ x2 ⭐ 10 0.100 ⭐ x3 ⭐ 10

(6.56)

0.100 ⭐ x4 ⭐ 10 where x ⫽ [h, l, t, b]T c3 ⫽ 0.10471 ($/in.3) c4 ⫽ 0.04811 ($/in.3) The lower bound on x1 is from the problem statement, whereas the lower bonds on x2, x3, and x4 are necessary to prevent selection of zero design variables, where functions in the constraint set are not defined. We prescale the constraints by dividing by suitable powers of 10 such that all constraint values are in the range [0, 10] at x (0) ⫽ [1, 7, 4, 2]T, to get the following function values:

292


ƒ(x (0)) ⫽ 15.81545 ($) g1(x (0)) ⫽ 8.6785 ⫻ 10⫺1 g2(x (0)) ⫽ 1.4250 ⫻ 10⫺1 g3(x (0)) ⫽ 1.0000 ⫻ 10⫺1 g4(x (0)) ⫽ 1.8319 ⫻ 10⫺1 g5(x (0)) ⫽ 2.3285 ⫻ 10⫺1 Since x (0) is feasible and K ⫽ 0, one unconstrained DFP search is completed before the scale factors are calculated. The new scale factors are given in Table 6.5, where the ␣j values scale the constraints and the ␩i values scale the variables. Table 6.6 gives the progress of the method at the end of each unconstrained DFP stage. At termination, x (T) ⫽ [0.2444, 6.2187, 8.2915, 0.2444]T and ƒ(x (T)) ⫽ 2.38116 ($) g1(x (T)) ⫽ ⫺6.5386 ⫻ 10⫺7 g2(x (T)) ⫽ ⫺1.5824 ⫻ 10⫺7 g3(x (T)) ⫽ ⫺2.6740 ⫻ 10⫺6 g4(x (T)) ⫽ 8.7002 ⫻ 10⫺8 g5(x (T)) ⫽ 2.324 ⫻ 10⫺1 The results indicate that the method found the vicinity of the solution quickly but did require additional iterations to focus the estimate of the solution before termination. Note that constraints 1–4 are tight at the solution, which seems logical. Even though only six iterations (stages) of the MOM algorithm are required, it should be noted in fairness that 1173 objective function and 1177 constraint set evaluations were required. The results given here are for a value of R ⫽ 20, and similar performance is obtained with values of R in the range Table 6.5

i, j ␣j ␩i

Welded Beam Scale Factors

1 0.8724 1.0000

2 0.5324 0.2704

3 0.9649 0.1441

4 1.6730 3.90

5 7.2657

6.4

Table 6.6

293

SUMMARY

BIAS Welded Beam Results (R ⴝ 20)

Stage (t)

x1

x2

x3

x4

ƒ(x)

P(x, ␴)

Cumulative DFP Iterations

1 2 3 4 5 6

0.2544 0.2567 0.2470 0.2450 0.2444 0.2444

5.7098 5.6319 6.1130 6.1937 6.2177 6.2187

8.3347 8.2979 8.2909 8.2913 8.2915 8.2915

0.2417 0.2435 0.2444 0.2444 0.2444 0.2444

2.3184 2.3182 2.3727 2.3792 2.3811 2.3811

2.3482 2.3476 2.3804 2.3810 2.3811 2.3811

27 36 44 64 71 76

[10, 100]. Typically, higher R values will produce fewer MOM stages with more DFP iterations per stage. The method will fail when values of R are too small. For instance, when R ⫽ 1, BIAS terminated after 14 stages and 57 DFP iterations. At termination, only two of the constraints were tight and x3 had not changed from x (0) 3 . This demonstrated the relationship between the scaling algorithm and the choice of R. 6.3.8

Other MOM-Type Codes

There are, of course, other good implementations of the MOM. In fact, Schittkowski [67], in his comparative experiments, tested no less than 12 MOM codes. We mention a few of these codes here. ACDPAC, by M. J. Best and A. T. Bowler [68], uses the Rockafellar [54] augmented Lagrangian form with special provision for variable bounds and linear constraints. The Best and Ritter [69] conjugate direction method is employed for the subproblems. GAPFPR, by D. M Himmelblau [70], uses the Hestenes [56] form, with inequalities transformed to equalities via slack variables. A quasi-Newton method with special provision for linear constraints is used for the subproblems. VF01A, by R. Fletcher, uses the Fletcher multiplier form [71] and a quasi-Newton method for the subproblems. P. E. Gill and W. Murray [72] have developed three related codes, SALQDR, SALQDF, and SALMNF, which employ an augmented Lagrangian form with inequalities transformed via slack variables. The various versions employ different unconstrained subproblem solution techniques and require either analytical or numerical derivatives. Additional data and discussion of the relative merits of these and other codes are given in Chapter 12.

6.4

SUMMARY

In this chapter we have examined transformation strategies with particular emphasis on augmented Lagrangian or multiplier methods. We have observed

294


that the standard penalty forms—the log, inverse, and bracket operator—are simple in concept and implementation. In addition we noted the inevitable ill-condition of the unconstrained subproblems associated with the infinitelimit value of the penalty parameter R and the resulting distortion of the penalty function contours. The MOM was examined as a remedy, and favorable computational characteristics were noted. Examples were given to support (and in some case significantly contribute to) the discussion. Easy-to-use transformation-based codes are now available and are discussed further in Chapter 12.

REFERENCES 1. Everett, H., ‘‘Generalized Lagrange Multiplier Method for Solving Problems of Optimum Allocation of Resources,’’ Oper. Res., 11, 300–471 (1963). 2. Siddall, J. N., Analytical Decision-Making in Engineering Design, Prentice-Hall, Englewood Cliffs, NJ, 1972. 3. Siddall, J. N., ‘‘Opti-Sep: Designer’s Optimization Subroutines,’’ McMaster University, Canada, 1971. 4. Lootsma, F. A., ‘‘A Survey of Methods for Solving Constrained Minimization Problems via Unconstrained Minimization,’’ in Numerical Methods for Nonlinear Optimization (F. A. Lootsma, Ed.), Academic, New York, 1972, pp. 313–348. 5. Ragsdell, K. M., R. R. Root, and G. A. Gabriele, ‘‘Subroutine ZERO,’’ in MELIB: A Subroutine Library for Mechanical Engineers, 1975. 6. Fiacco, A. V., and G. P. McCormick, ‘‘Computational Algorithm for the Sequential Unconstrained Minimization Technique for Nonlinear Programming,’’ Manag. Sci., 10, 601–617 (1964). 7. Fiacco, A. V., and G. P. McCormick, Nonlinear Programming: Sequential Unconstrained Minimization Techniques, Wiley, New York, 1968. 8. Lootsma, F. A., ‘‘Extrapolation in Logarithmic Programming,’’ Philips Res. Rep., 23, 108–116 (1968). 9. Lootsma, F. A., ‘‘Boundary Properties of Penalty Functions for Constrained Minimization,’’ Philips Res. Rep. Suppl., 3, 1970. 10. Fletcher, R., and A. P. McCann, ‘‘Acceleration Techniques for Nonlinear Programming,’’ in Optimization (R. Fletcher, Ed.), Academic, New York, 1969, pp. 203– 214. 11. Murray, W., ‘‘Ill-Conditioning in Barrier and Penalty Functions Arising in Constrained Nonlinear Programming,’’ in Proceedings of the Sixth Symposium on Mathematical Programming, Princeton, NJ, 1967. 12. Lootsma, F. A., ‘‘Hessian Matrices of Penalty Functions for Solving Constrained Optimization Problems,’’ Philips Res. Rep., 24, 322–331 (1969). 13. Courant, R., ‘‘Variational Methods for the Solution of Problems of Equilibrium and Vibrations,’’ Bull. Am. Math. Soc., 49, 1–23 (1943). 14. Carroll, C. W., ‘‘The Created Response Surface Technique for Optimizing Nonlinear Restrained Systems,’’ Oper. Res., 9, 169–184 (1961).

REFERENCES

295

15. Fiacco, A. V., and G. P. McCormick, ‘‘The Sequential Unconstrained Minimization Technique for Nonlinear Programming, a Primal-Duel Method,’’ Manag. Sci., 10, 360–366 (1964). 16. Fiacco, A. V., and G. P. McCormick, ‘‘Computational Algorithm for the Sequential Unconstrained Minimization Technique for Nonlinear Programming,’’ Manag. Sci., 10, 601–617 (1964). 17. Fiacco, A. V., and G. P. McCormick, Nonlinear Programming: Sequential Unconstrained Minimization Techniques, Wiley, New York, 1968. 18. Zangwill, W. I., ‘‘Nonlinear Programming via Penalty Functions,’’ Manag. Sci., 13(5), 344–358 (1967). 19. Lootsma, F. A., ‘‘Logarithmic Programming: A Method of Solving Nonlinear Programming Problems,’’ Philips Res. Rep., 22, 329–344 (1967). 20. Lootsma, F. A., ‘‘Constrained Optimization via Penalty Functions,’’ Philips Res. Rep., 23, 408–423 (1968). 21. Lootsma, F. A., ‘‘A Survey of Methods for Solving Constrained-Minimization Problems via Unconstrained Minimization,’’ in Optimization and Design (M. Avriel, H. Rijckaert, and D. J. Wilde, Eds.), Prentice-Hall, Englewood Cliffs, NJ, 1973, pp. 88–118. 22. McCormick, G. P., ‘‘Penalty Function versus Non-Penalty Function Methods for Constrained Nonlinear Programming Problems,’’ Math. Prog., 1(2), 217–238 (1971). 23. Lasdon, L. S., ‘‘An Efficient Algorithm for Minimizing Barrier and Penalty Functions,’’ Math. Prog., 2(1), 65–106 (1972). 24. Moe, J., ‘‘Penalty-Function Methods,’’ in Optimum Structural Design: Theory and Applications (R. H. Gallagher and O. C. Zienkiewicz, Eds.), Wiley, New York, 1973, Chap. 9. 25. Ryan, D. M., ‘‘Penalty and Barrier Functions,’’ in Numerical Methods for Constrained Optimization (P. E. Gill and W. Murray, Eds.), Academic, New York, 1974, pp. 175–190. 26. Fox, R. L., Optimization Methods for Engineering Design, Addison-Wesley, Reading, MA, 1971. 27. Himmelblau, D. M., Applied Nonlinear Programming, McGraw-Hill, New York, 1972, Chap. 7, p. 299. 28. Avriel, M., Nonlinear Programming: Analysis and Methods, Prentice-Hall, Englewood Cliffs, NJ, 1976. 29. Gulf Oil Corporation, ‘‘Compute II, a General Purpose Optimizer for Continuous Nonlinear Models—Description and User’s Manual,’’ Gulf Computer Sciences Incorporated, Houston, TX, 1972. 30. Afimiwala, K. A., ‘‘Program Package for Design Optimization,’’ Department of Mechanical Engineering, State University of New York at Buffalo, 1974. 31. Lill, S., The University of Liverpool Computer Laboratory: E04HAF, Numerical Algorithms Group, Document No. 737, 1974. 32. DeSilva, B. M. E., and G. N. C. Grant, ‘‘Comparison of Some Penalty Function Based Optimization Procedures for the Synthesis of a Planar Truss,’’ Int. J. Numer. Methods Eng., 7(2), 155–173 (1973). 33. Bracken, J., and G. P. McCormick, Selected Applications of Nonlinear Programming, Wiley, New York, 1968.

296


34. Gupta, B. D., and S. S. Rao, ‘‘Optimum Design of Air-Conditioned Buildings,’’ ASME J. Mech. Des., 102(3), 476–480 (1980). 35. Rao, S. S., and B. D. Gupta, ‘‘Application of External Distributions in the Design of Thermal Systems,’’ ASME J. Mech. Des., 102(3), 481–491 (1980). 36. Fiacco, A. V., and G. P. McCormick, Nonlinear Programming, Wiley, New York, 1968. 37. Sandgren, E., ‘‘The Utility of Nonlinear Programming Algorithms,’’ Ph.D. dissertation, Purdue University, 1977; available from University Microfilm, 300 N. Zeeb Road, Ann Arbor, MI, Document No. 7813115. 38. Fletcher, R., Practical Methods of Optimization, 2nd ed., Wiley, New York, 2004. 39. Nocedal, J., and S. Wright, Numerical Optimization, Springer Science, New York, 1999. 40. Franz, J., M. Liepelt, and K. Schittkowski, ‘‘Penalty-Barrier-Methods for Nonlinear Optimization: Implementation and Numerical Results,’’ Report, Department of Mathematics, University of Bayreuth, Germany. 41. http: / / www / unibayreuth.de / departments / math / ⬃kschittkowski / software.htm. 42. http: / / www.ece.northwestern.edu / OTC / . 43. Gabriele, G. A., OPTLIB: An Optimization Program Library, Purdue University, West Lafayette, IN, 1979. 44. Bertsekas, D. P., Constrained Optimization and Lagrange Multiplier Methods, Athena Scientific, Belmont, MA, 1996. 45. Kowalik, J., ‘‘Nonlinear Programming Procedures and Design Optimization,’’ Acta Polytech. Scand., 13, 1966. 46. Fiacco, A. V., and G. P McCormick, ‘‘The Sequential Unconstrained Minimization Technique without Parameters,’’ Oper. Res., 15, 820–827 (1967). 47. Fiacco, A. V., ‘‘Sequential Unconstrained Minimization Methods for Nonlinear Programming,’’ Thesis, Northwestern University, Evanston, IL, 1967. 48. Rosenbrock, H. H., ‘‘An Automatic Method for Finding the Greatest or Least Value of a Function,’’ Comput. J. 3, 175–184 (1960). 49. Huard, P., ‘‘Resolution of Mathematical Programming with Nonlinear Constraints by the Method of Centres,’’ in Nonlinear Programming (J. Abodie, Ed.), NorthHolland, Amsterdam, 1967, pp. 207–219. 50. Staha, R. L., ‘‘Documentation for Program Comet, a Constrained Optimization Code,’’ University of Texas, Austin, TX, 1973. 51. Lootsma, F. A., ‘‘Convergence Rates of Quadratic Exterior Penalty-Function Methods for Solving Constrained Minimization Problems,’’ Philips Res. Rep., 29(1) (Feb. 1974). 52. Schuldt, S. B., ‘‘A Method of Multipliers for Mathematical Programming Problems with Equality and Inequality Constraints,’’ JOTA, 17(1 / 2), 155–162 (1975). 53. Schuldt, S. B., G. A. Gabriele, R. R. Root, E. Sandgren, and K. M. Ragsdell, ‘‘Application of a New Penalty Function Method to Design Optimization,’’ ASME J. Eng. Ind., 99(1), 31–36 (Feb. 1977). 54. Rockafellar, R. T., ‘‘A Dual Approach to Solving Nonlinear Programming Problems by Unconstrained Optimization,’’ Math. Prog., 5(3), 354–373 (1973). 55. Powell, M. J. D., ‘‘A Method for Nonlinear Constraints in Minimization Problems,’’ in Optimization (R. Fletcher, Ed.), Academic, London, 1969, pp. 283–298.

REFERENCES

297

56. Hestenes, M. R., ‘‘Multiplier and Gradient Methods,’’ JOTA, 4(5), 303–320 (1969). 57. Gould, F. J., and S. Howe, ‘‘A New Result on Interpreting Lagrange Multipliers as Dual Variables,’’ Department of Statistics, University of North Carolina, Chapel Hill, NC, 1971. 58. Gill, P. E., and W. Murray, ‘‘The Computation of Lagrange-Multiplier Estimates for Constrained Minimization,’’ Math. Prog., 17(1), 32–60 (1979). 59. Root, R. R., ‘‘An Investigation of the Method of Multipliers,’’ Ph.D. dissertations, Purdue University, West Lafayette, IN, Dec. 1977. 60. Bertsekas, D. P., ‘‘On Penalty and Multiplier Methods for Constrained Minimization,’’ in Nonlinear Programming, Vol. 2 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, Eds.), Academic, New York 1975, pp. 165–192. 61. Kort, B. W., ‘‘Rate of Convergence of the Method of Multipliers with Inexact Minimization,’’ in Nonlinear Programming, Vol. 2 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, Eds.), Academic, New York, 1975, pp. 193–214. 62. Root, R. R., and K. M. Ragsdell, ‘‘Computational Enhancements of the Method of Multipliers,’’ ASME J. Mech. Des., 102(3), 517–523 (1980). 63. Keefer, D. L., and X. X. Gottfried, ‘‘Differential Constraint Scaling in Penalty Function Optimization,’’ AIIE Trans., 2(4), 281 (1970). 64. Keefer, D. L., ‘‘Simpat: A Self-Bounding Direct Search Method for Optimization,’’ I & EC Proc. Des. Develop., 12, 92 (1973). 65. Ragsdell, K. M., and D. T. Phillips, ‘‘Optimal Design of a Class of Welded Structures Using Geometric Programming,’’ ASME J. Eng. Ind. Ser. B, 98(3), 1021– 1025 (1976). 66. Root, R. R., and K. M. Ragsdell, ‘‘BIAS: A Nonlinear Programming Code in Fortran IV-Users Manual,’’ Purdue Research Foundation, West Lafayette, IN, June 1976. 67. Schittkowski, K., ‘‘Nonlinear Programming Codes: Information, Tests, Performance,’’ Lecture Notes in Economics and Mathematical Systems, Vol. 183, SpringerVerlag, Berlin, 1980. 68. Best, M. J., and A. T. Bowler, ‘‘ACDPAC: A Fortran IV Subroutine to Solve Differentiable Mathematical Programmes—Users Guide—Level 2.0,’’ Res. Rep., CORR75-26, Department of Combinatorics and Optimization, University of Waterloo, Ontario, Canada, 1978. 69. Best, M. J., and Ritter, K., ‘‘An Accelerated Conjugate Direction Method to Solve Linearly Constrained Minimization Problems,’’ J. Computer Syst. Sci., 11(3), 295– 322 (1975). 70. Newell, J. S., and D. M. Himmelblau, ‘‘A New Method for Nonlinearly Constrained Optimization,’’ AIChE J., 21(3), 479–486 (1975). 71. Fletcher, R., ‘‘An Ideal Penalty Function for Constrained Optimization,’’ in Nonlinear Programming, Vol. 2 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, Eds.), Academic, New York, 1975. 72. Gill, P. E., and W. Murray (Eds.), Numerical Methods for Constrained Optimization, Academic, New York, 1974. 73. Freudenstein, F., ‘‘Approximate Synthesis of Four-Bar Linkages,’’ Trans. ASME, 853–861 (Aug. 1955).

298


PROBLEMS

6.1. Describe briefly the major difference between the MOM and other transformation methods such as SUMT. 6.2. Under what conditions does the MOM produce no distortion in the curvature of the penalty function contours from stage to stage? 6.3. Would it be possible and/or desirable to devise an interior version of the MOM? Why or why not? 6.4. For the problem Minimize

ƒ(x)


j ⫽ 1, 2, 3, . . . , J

we propose to use the two penalty functions

冘具g (x)典 J

P1(x, R) ⫽ ƒ(x) ⫹ R

2

j

j⫽1

where R is positive and increases in magnitude at the end of each unconstrained subproblem, and

冘具g (x) ⫹ ␴ 典 ⫺ R 冘 ␴ J

P2(x, R, ␴) ⫽ ƒ(x) ⫹ R

J

2

j

j

j⫽1

2 j

j⫽1

with R fixed and ␴j adjusted after each unconstrained subproblem to force convergence. Discuss the relative merits of the two penalty forms, including any computational differences. 6.5. Consider the problem Minimize

ƒ(x) ⫽ 100(x2 ⫺ x21)2 ⫹ (1 ⫺ x1)2

Subject to g1(x) ⫽ x1 ⫹ 1 ⭓ 0 g2(x) ⫽ 1 ⫺ x2 ⭓ 0 g3(x) ⫽ 4x2 ⫺ x1 ⫺ 1 ⭓ 0 g2(x) ⫽ 1 ⫺ 0.5x1 ⫺ x2 ⭓ 0 Construct P using the log penalty, inverse penalty, and MOM penalty forms. Let x (0) ⫽ [⫺1, 1]T and [⫺0.5, 0.5]T in each case. Discuss any difficulties encountered with these penalty forms at x (0).

PROBLEMS

299

6.6. For the inequality-constrained problem [gj (x) ⭓ 0, j ⫽ 1, 2, 3, . . . , J], suppose we start with an interior point. Explain whether the following functions act as interior or exterior penalty methods and why.

冘 {g (x)} J

(a) P(x, R) ⫽ ƒ(x) ⫹ R

j

⫺2

j⫽1

(b) P(x, R) ⫽ ƒ(x) ⫹ R[min{0, min [gj (x)]}2] j

6.7. Consider the problem Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ (x2 ⫺ 1)2

Subject to g1(x) ⫽ x1 ⫹ x2 ⫺ 1 ⭓ 0 g2(x) ⫽ 1 ⫺ x1 ⭓ 0 g3(x) ⫽ 1 ⫺ x2 ⭓ 0 and the penalty function

冘冋g 1(x)册 J

P(x, R) ⫽ ƒ(x) ⫹ R

j⫽1

j

(a) Find the stationary points of P(x, R) as a function of R. (b) Demonstrate the influence of g2 and g3 on the convergence rate of the method. (c) Will the method converge more or less quickly when g2 and g3 are ignored? 6.8. Find the solution to the following problem: Minimize

ƒ(x) ⫽ x21 ⫹ x22 ⫹ x23

Subject to g1(x) ⫽ 1 ⫺ x2⫺1x3 ⭓ 0 g2(x) ⫽ x1 ⫺ x3 ⭓ 0 h1(x) ⫽ x1 ⫺ x22 ⫹ x2x3 ⫺ 4 ⫽ 0 0 ⭐ x1 ⭐ 5 0 ⭐ x2 ⭐ 3 0 ⭐ x3 ⭐ 3 using an appropriate transformation method of your choice. Find x to three decimal places. 6.9. We know that it is possible to define ‘‘curvature’’ K and ‘‘radius of curvature’’ R at a point of a function. Let the function be a particular

300


contour value ƒ of ƒ(x) in two dimensions. Develop a relationship between ⵜ2ƒ(x) and K. 6.10. Discuss how one might use the relation in problem 6.9 to rate the quality of the various penalty forms. 6.11. Minimize-Perimeter Enclosure Problem. Consider the three circular plane objects in Figure 6.23. We wish to arrange the objects in the positive quadrant such that the circumscribed rectangle has a perimeter of minimum length. We could obviously extend the formulations to include objects of various shapes and numbers. The problem has a pleasing simplicity but is of practical value. One application is the planar packaging of electronic components. Another is the packing of a template for mass producing stamped metal parts. Let the design variables be the (x, y) locations of the object centers. The objective is the enclosure perimeter, ƒ(x) ⫽ 2(A ⫹ B), where A and B are the lengths of the rectangle sides. Force the objects to stay in the positive quadrant by writing constraints of the form gj (x) ⫽ xa ⫺ Ra ⭓ 0. To avoid object interference, insist that the distance between two objects is greater than or equal to the sum of their radii: gj (x) ⫽ {(xa ⫺ xb)2 ⫹ (ya ⫺ yb)2}1 / 2 ⫺ (Ra ⫹ Rb) ⭓ 0 Assume that the circle radii are given as shown in Figure 6.24.

Figure 6.23. Minimum-perimeter enclosure problem.

PROBLEMS

301

Figure 6.24. R ⫽ 5, 10, 15.

(a) Formulate the complete NLP. (b) Find the minimum-perimeter enclosure using a transformation method of your choice. (c) Is the feasible region convex? (d) What is the likelihood of encountering local solutions? (e) Check your solution to (b) using a piece of paper, circle cutouts, a ruler, and experimentation by hand. 6.12. Mechanism Precision Point Placement. Consider the four-bar mechanism shown in Figure 6.25. Using this planar device, we wish to coordinate the angles ␪ and ␾ such that the desired function yd(x) is approximated. In particular, we wish to locate the three precision points x1, x2, x3 such that the square of the error in the approximation is least. Consider the generated and desired functions in Figure 6.26 and the resulting error function in Figure 6.27, where E(x) ⫽ yd(x) ⫺ yg(x). The input–output relationship is constructed given xL, xH, yL, yH and Freudenstein’s equation [73]: F(␾) ⫽ R1 cos ␪ ⫺ R2 cos ␾ ⫹ R3 ⫺ cos(␪ ⫺ ␾) ⫽ 0 where R1 ⫽ d/c, R2 ⫽ d/a, R3 ⫽ (a2 ⫺ b2 ⫹ c2 ⫹ d2)/2ca, by setting E(x) ⫽ 0 at x1, x2, and x3. In addition, it is assumed that the linear relations ␪ ⫽ c1x ⫹ c2 and ␾ ⫽ c3y ⫹ c4 are known.

302


Figure 6.25. Angle coordination synthesis problem.

(a) Formulate an appropriate NLP for precision point placement. (b) Let yd(x) ⫽ 1/x, xL ⫽ 1, xH ⫽ 2, yL ⫽ 1, yJ ⫽ 0.5, so that ␪ ⫽ 90x ⫹ 60 and ␾ ⫽ ⫺180y ⫹ 240; find the precision points x1, x2, x3 given x (0) ⫽ [1.15, 1.50, 1.85]T. 6.13. Consider the problem Minimize

ƒ(x) ⫽ (x ⫺ 2)2

x 僆 R1

Subject to h(x) ⫽ x ⫺ 3 ⫽ 0 shown in Figure 6.28. Obviously the solution is x* ⫽ 3 with ƒ(x*) ⫽

Figure 6.26. Precision point placement.

PROBLEMS

303

Figure 6.27. Output error function.

1.0. Develop relations for (a) the parabolic penalty and (b) the MOM penalty forms that support the contour plots given in Figures 6.29 and 6.30. 6.14. Solve problem 3.19 (without using the volume constraint to eliminate one of the design variables), using a penalty method of your choice. 6.15. (a) Solve problem 3.20 (without using the flow-rate constraint to eliminate one of the design variables), using the method of multipliers. (b) Discuss the difficulties associated with using a gradient-based method for the subproblems.

Figure 6.28. Single-variable constrained problem.

304


Figure 6.29. Parabolic penalty contours.

Figure 6.30. MOM contours.

7 CONSTRAINED DIRECT SEARCH The transformation- and Lagrangian-based methods discussed in Chapter 6 offer implicit ways of handling constraints that avoid consideration of the detailed structure of the constraint functions. The sole structural distinction is between equality and inequality constraints. By contrast, the methods to be considered in this and the next several chapters operate on the constraints explicitly, that is, they directly take into account the constraints in the course of the optimization iterations. In this chapter we focus on direct-search methods, that is, techniques for locating constrained optima that rely upon only objective and constraint function values to guide the search. As in the unconstrained case, the motivation for considering such methods is that in engineering applications it is frequently necessary to solve problems whose functions are discontinuous or nondifferentiable, and in such cases only function-value-based approaches are applicable. We consider two families of methods: those that are essentially adaptations of the unconstrained direct-search methods discussed in Chapter 3 and those that rely upon the random selection of trail points. Both families of methods are heuristic in nature, that is, they are based on intuitive constructions unsupported by rigorous theory and thus offer no guarantees of convergence or estimates of rates of convergence. In spite of these limitations, we find these techniques to be quite reliable in yielding improved if not optimal solutions and easy to implement because of their simple search logic. We begin with a brief discussion of the steps required to prepare general constrained problems for solution via direct-search methods.


305

306

CONSTRAINED DIRECT SEARCH

7.1

PROBLEM PREPARATION

The general constrained optimization problem involves inequality constraints, equality constraints, and variable upper and lower bounds. Intuitively, given a feasible point x 0, it should be possible to modify the function-value-based search logic of any unconstrained method to check that the trial points generated with x 0 as base also satisfy the problem constraints. For instance, if v is a trial point that violates some inequality constraint or variable bound, then modify v by reducing the step take from x 0, so that these constraints are again satisfied. Equality constraints, however, are more troublesome, because even if x 0 and v satisfy the equalities, the points on the line x ⫽ x 0 ⫹ ␣(v ⫺ x 0)

for 0 ⭐ ␣ ⭐ 1

need not satisfy them, except in the case of linear equalities. If in addition v is infeasible, then it is difficult to readjust the components of v to ensure satisfaction of nonlinear equality constraints using only the constraint function values. It is, in fact, equivalent to solving the simultaneous equation set hk(x) ⫽ 0

k ⫽ 1, . . . , K

given the initial estimate v using only values of the hk(x) functions and adjusting the components of v so that the resulting trial point is an improvement over the base point x 0. As a result of the difficulty of ensuring the satisfaction of equality constraints, virtually all direct-search methods require that problems be posed only in terms of inequality constraints. Equality constraints must be eliminated from the problem explicitly or implicitly prior to solution. A second, even more basic, problem is that direct-search methods must typically be initiated with a feasible point. To be sure, in many engineering applications a feasible starting point can be defined on the basis of prior knowledge about the system or based on the operating conditions of an existing system. However, this is not always the case, and thus systematic procedures to locate a feasible point are a matter of practical concern.

7.1.1

Treatment of Equality Constraints

The simplest approach to eliminating equality constraints is to solve each for one variable and use the resulting expression to eliminate that variable from the problem formulation by substitution. In doing so, however, care must be exercised to ensure that any specified bounds on the substituted variable are retained as inequality constraints.

7.1

PROBLEM PREPARATION

307

Example 7.1

Consider the problem Minimize

ƒ(x) ⫽ x21 ⫹ 4x22 ⫹ x23

Subject to h1(x) ⫽ x1 ⫹ x22 ⫺ 2 ⫽ 0 ⫺1 ⭐ x1 ⭐ 1

0 ⭐ (x2, x3) ⭐ 2 Clearly the constraint h1 can be solved for x1 to yield x1 ⫽ 2 ⫺ x22, and thus the objective function can be reduced by substitution to ƒ(x2, x3) ⫽ (2 ⫺ x22)2 ⫹ 4x22 ⫹ x32 To ensure satisfaction of the bounds on x1, two inequality constraints must be imposed: ⫺1 ⭐ 2 ⫺ x22 ⭐ 1

The resulting reduced-dimensionality problem becomes Minimize

(2 ⫺ x22)2 ⫹ 4x22 ⫹ x23

Subject to

1 ⭐ x2 ⭐ 兹3

0 ⭐ x3 ⭐ 2

Note that if the bounds on x1 had been ignored, the minimum solution x2

⫽ x3 ⫽ 0 of the reduced-dimensionality problem would lead to an infeasible x1, namely, x1 ⫽ 2.

In implementing the equality constraint elimination strategy, it is normally neither necessary nor desirable to actually carry out algebraic substitution. Instead, it is more convenient and much more reliable to simply include the expressions obtained by solving a constraint for a variable as statements defining the value of that variable in the routine for evaluating the problem’s objective function and inequality constraints. In this way, values of the dependent variable can be saved for subsequent printing, and the possibility of algebraic substitution errors is minimized. While substitution is an obvious way to eliminate equality constraints, it is not always possible to carry it out explicitly. In many cases it is either algebraically infeasible or inconvenient to actually solve an equation for one variable in terms of the others. In those cases, implicit elimination is possible by numerically solving the constraint for the value of the dependent variable

308


corresponding to each selected trial value of the independent optimization variables. Such numerical treatment of equality constraints imposes a large computational overhead, particularly if the number of dependent variables is large, but must be accepted if direct-search methods are to be employed. Example 7.2

The problem Minimize

ƒ(x) ⫽ x1x2x3

Subject to h1(x) ⫽ x1 ⫹ x2 ⫹ x3 ⫺ 1 ⫽ 0 h2(x) ⫽ x21x3 ⫹ x2x23 ⫹ x2⫺1x1 ⫺ 2 ⫽ 0 0 ⭐ (x1, x3) ⭐ –21 involves two equality constraints and can therefore be viewed as consisting of two dependent variables and one independent variable. Clearly, h1 can be solved for x1 to yield x1 ⫽ 1 ⫺ x2 ⫺ x3 Thus, upon substitution the problem reduces to Minimize

(1 ⫺ x2 ⫺ x3)x2x3

Subject to

(1 ⫺ x2 ⫺ x3)2x3 ⫹ x2x23 ⫹ x⫺1 2 (1 ⫺ x2 ⫺ x3) ⫺ 2 ⫽ 0 0 ⭐ 1 ⫺ x2 ⫺ x3 ⭐ –21 0 ⭐ x3 ⭐ –21

It is very difficult to solve the remaining quality constraint for one variable, say x3, in terms of the other. Instead, for each value of the independent variable x2, the corresponding value of x3 must be calculated numerically using some root-finding method. Note that in this case the bounds and inequalities involving x3 become implicit functions of the independent variable x2. That is, since x3 and x2 are related through some function q(x2) ⫽ x3, the constraints become 0 ⭐ 1 ⫺ x2 ⫺ q(x2) ⭐ –12 0 ⭐ q(x2) ⭐ –12 which can be evaluated only when the equality constraint is solved.

7.2

ADAPTATIONS OF UNCONSTRAINED SEARCH METHODS

309

In general, if the number of equality constraints is large, then it becomes appropriate to use equation-ordering and decomposition procedures to automate the selection of independent variables and the sequencing of the equality constraints for efficient solution. The discussion of these techniques is beyond the scope of this book. The reader is directed to texts dealing with the solution of equation sets [1, 2] or the excellent review given by Sargent [3]. 7.1.2

Generation of Feasible Starting Points

Borrowing from the terminology introduced in the linear programming context, the process of generating a feasible starting point is often called a phase I procedure. Three such procedures are in common use: random generation, minimization of an unconstrained penalty-type function, and sequential constrained minimization. In this section we review briefly only the first type of strategy. The other two are considered in Chapter 13. Conceptually the simplest way of generating feasible starting points is to use pseudo-random-number generation routines (available on most computer systems) to generate trial points, which are then tested for feasibility by evaluating the constraint functions. Specifically, the selected variable upper and lower bounds are used to calculate the components xi ⫽ xi(L) ⫹ ri(xi(U) ⫺ xi(L))

i ⫽ 1, . . . , N

(7.1)

where ri is a random number distributed uniformly on the interval (0, 1). If x is an N-component vector, then ri can be sampled N times and the xi calculated from (7.1). This point is then used to evaluate the problem constraint functions, and the process is continued until a point satisfying all constraints is found. Clearly, since the probability of finding a point simultaneously satisfying a set of equality constraints in this fashion is quite low, it is appropriate to use the form of the problem in which the quality constraints have been eliminated. The sampling is thus done only on the independent variables. Even in this case, random generation can require extensive computation time. Thus, for higher dimensionality and tightly constrained problems, the alternate strategies discussed in Chapter 13 may be necessary. 7.2


In this section we briefly review the difficulties that can arise in attempting to adapt the direct-search methods of Chapter 3 to constrained problems. Throughout this discussion it is assumed that any equality constraints have been eliminated using the devices of Section 7.1.1. We consider in detail only one of the direct-search adaptations, the complex method, since this algorithm has enjoyed wide use in engineering applications.

310


7.2.1

Difficulties in Accommodating Constraints

In principle, all three of the basic direct-search methods considered in Chapter 3 could be modified to accommodate inequality constraints. In essence, before accepting any candidate point, the constraints could be checked, and if they are violated, the current point could be rejected and another generated that would satisfy the constraints. In practice, however, considerable care must be exercised in defining the rules according to which infeasible points are modified to correct for infeasibility. Otherwise, the resulting algorithm not only may suffer from a substantially reduced rate of convergence but also may terminate prematurely. Let us first consider the adaptation of the conjugate directions method. Given an initial feasible point x 0 and a set of search directions di, i ⫽ 1, . . . , N, the search would proceed with line searches along each of these direction vectors so as to both generate conjugate directions and improve the objective function value. In the presence of inequality constraints, any given line search may well lead to a one-parameter optimum that is infeasible. Clearly, to retain feasibility the search along the direction vector must be terminated short of the one-parameter optimum at the point at which the boundary of the feasible region is intersected. In this case, however, the generation of conjugate directions will no longer be guaranteed, and thus the search can degenerate to a series of line searches along a fixed set of direction vectors. As shown in Figure 7.1, this in turn can result in premature termination. In the figure, the line searches in the d1 and d2 directions from the point x(t) either lead outside the feasible region or to points with worse objective function values. Consequently, the search will terminate at x(t). Evidently, one way to escape the dilemma is to redefine the set of search directions whenever further progress is stymied. For instance, a new direction vector parallel to constraint g2 would lead to improvement. Constructions of this nature have been proposed by Davies (cited in Chap. 5 of ref. 4,) and others, but efficient implementations typically require the use of constraint gradient information. Similar modifications have been proposed for use with the pattern search method. Recall that in this algorithm progress is made either through an exploratory move or through a pattern move. In the presence of constraints, either or both of these moves may lead to infeasible points, and thus some corrective rules must be added to accommodate these situations. As shown in Figure 7.2, if a pattern move results in an infeasible tentative base point, then the step can simply be reduced by a fixed fraction, say –12 , until a feasible point is found. Similarly, during exploration, infeasible points can simply be treated as points with a very high objective function value and thus rejected as shown in Figure 7.3. Eventually, a point will be reached at which all exploratory steps lead to either infeasible or feasible but worse points. In this case, the step sizes used in exploration can be reduced uniformly and the exploratory move restarted. However, as in the conjugate directions case (see

7.2


Figure 7.1. Premature termination of conjugate directions adaptation.

Figure 7.2. Retraction of pattern step.

311

312


Figure 7.3. Rejection of infeasible exploratory step.

Figure 7.1), it is nonetheless possible to reach base points for which step-size reduction does not lead to improvement. Again, this situation can be rectified only by developing search directions that lie parallel to the constraint surface. Proposals that accomplish this have been made by Klingman and Himmelblau [5] and others (see the discussion in ref. 6, Chap. 7, as well as ref. 4). However, in all cases constraint and objective function gradient information is required to carry out the proposed constructions. If gradient values are available, then, of course, experience has shown that the methods of Chapters 8– 10 are much to be preferred. The conclusions we can draw from this brief consideration of direct search in the presence of constraints is that the simple device of retracting a step if it leads to infeasibility is an adequate way of ensuring trial-point feasibility. This device, of course, requires that the search be begun with a feasible point and that successive base points remain feasible so that retraction is always possible. On the other hand, the use of any set of search directions that does not adjust itself to allow movement along constraint surfaces is quite clearly unsatisfactory. This problem can, however, be mitigated if the search directions are widely scattered or in same sense randomized. This is effectively what is achieved in the adaptation of the simplex direct-search method to accommodate inequality constraints, which is discussed in the next section. 7.2.2

Complex Method

The simplex direct-search method is based on the generation and maintenance of a pattern of search points and the use of projections of undesirable points through the centroid of the remaining points as the means of finding new trial points. In the presence of inequality constraints, it is evident that if the new point is infeasible, then it is a simple matter to retract it toward the centroid until it becomes feasible. Clearly, this will immediately destroy the regularity of the simplex, but, as was shown by Nelder and Meade, this is not necessarily a bad result. However, the generation of the initial simplex of points is prob-

7.2


313

lematic, because for any given simplex size parameter ␣ it is highly likely that many of the points of the regular simplex will be infeasible. Evidently, as each point is calculated, it must be tested for feasibility and, if infeasible, suitably adjusted. As a consequence of this adjustment, the regularity of the simplex construction is again destroyed. Moreover, the points must be generated sequentially rather than being essentially simultaneously defined using the formula for a regular simplex. These considerations have led to the simplex modification proposed by Box [7] and called the complex method. Box proposed that the set of P trial points be generated randomly and sequentially. The mechanics of random point generation proceeds as in Section 7.1.2. Given variable upper and lower bounds x(U) and x(L), the pseudorandom variable uniformly distributed on the interval (0, 1) is sampled, and the point coordinates calculated via Eq. (7.1). Here N samples are required to define a point in N dimensions. Each newly generated point is tested for feasibility, and if infeasible it is retracted toward the centroid of the previously generated points until it becomes feasible. The total number of points to be used, P, should, of course, be no less than N ⫹ 1 but can be larger. Given this set of points, the objective function is evaluated at each point, and the point corresponding to the highest value is rejected. A new point is generated by reflecting the rejected point a certain distance through the centroid of the remaining points. Thus, if x R is the rejected point and x is the centroid of the remaining points, then the new point is calculated via x m ⫽ x ⫹ ␣(x ⫺ x R) The parameter ␣ determines the distance of the reflection: ␣ ⫽ 1 corresponds to setting the distance 㛳x m ⫺ x㛳 equal to 㛳x ⫺ x R㛳; ␣ ⬎ 1 corresponds to an expansion; ␣ ⬍ 1 to a contraction. Now, at the new point, the performance function and the constraints are evaluated. There are several alternatives: 1. The new point is feasible and its function value is not the highest of the set of points. In this case, select the point that does correspond to the highest and continue with a reflection. 2. The new point is feasible and its function value is the highest of the current set of P points. Rather than reflecting back again (which could cause cycling), retract the point by half the distance to the previously calculated centroid. 3. The new point is infeasible. Retract the point by half the distance to the previously calculated centroid. The search is terminated when the pattern of points has shrunk so that the points are sufficiently close together and/or when the differences between the function values at the points become small enough.

314


Box [7] has performed numerical experiments with this algorithm and on this empirical basis recommends using ␣ ⫽ 1.3 and P ⬇ 2N. Biles [8], on the other hand, reports good results with P ⫽ N ⫹ 2. Most implementations follow Box’s recommendation that if a point exceeds one of the bounds its corresponding coordinate be set equal to the bound. The ␣ ⬎ 1 compensates for the shrinking of the complex caused by halving the distances, while the large number of vertices is intended to prevent the complex from collapsing and flattening out along a constraint. The setting at the value of the bound is intended to avoid unnecessary contraction of the complex. The complete algorithm can be summarized as follows. Complex Method. Given an initial strictly feasible point x 0, reflection parameter ␣, and termination parameters ␧ and ␦:

Step 1. Generate the initial set of P feasible points. For each point p ⫽ 1, . . . , P ⫺ 1: (a) Sample N times to determine the point x pi . (b) If x p is infeasible, calculate the centroid x of the current set of points and reset x p ⫽ x p ⫹ –12 (x ⫺ x p) Repeat until x p becomes feasible. (c) If x p is feasible, continue with (a) until P points are available. (d) Evaluate ƒ(x p), for p ⫽ 0, . . . , P ⫺ 1. Step 2. Carry out the reflection step. (a) Select the point x R such that ƒ(x R) ⫽ max ƒ(x p) ⭤ Fmax (b) Calculate the centroid x and the new point x m ⫽ x ⫹ ␣(x ⫺ x R) (c) If x m is feasible and ƒ(x m) ⭓ Fmax, retract half the distance to the centroid x. Continue until ƒ(x m) ⬍ Fmax.. (d) If x m is feasible and ƒ(x m) ⬍ Fmax, go to step 4. (e) If x m is infeasible, go to step 3. Step 3. Adjust for feasibility. (a) Reset violated variable bounds: m (L) If x mi ⬍ x(L) i , set x i ⫽ xi . m (U) m If x i ⬎ xi , set x i ⫽ xi(U).

7.2


315

(b) If the resulting x m is infeasible, retract half the distance to the centroid. Continue until x m is feasible, then go to step 2(c). Step 4. Check for termination. (a) Calculate ƒ⫽

1 P

冘 ƒ(x ) p

x⫽

1 P

冘x

p

(b) If

冘 (ƒ(x ) ⫺ ƒ) ⭐ ␧ p

2

and

p

冘㛳(x

p

⫺ x)2兩 ⭐ ␦

p

terminate. Otherwise, go to step 2(a). Note that in order to carry out step 1, upper and lower bounds must be set on all variables. If such bounds do not naturally arise as part of the problem formulation, then it is necessary to provide some estimated values. These estimates must include the problem feasible region, and for computational efficiency the bounds should be as tight as possible. The detailed calculations of the complex method are illustrated in the next example. Example 7.3

A rectangular structure with an open frontal face is to be designed to enclose a plant unit. The structure is to have a volume of 16,000 ft3, but the perimeter of the base should be no more than 220 ft. The depth is to be no more than 60 ft and the width no more than 80 ft. Moreover, the width should not exceed three times the depth, and the height is to be no more than two-thirds of the width. The cost of the corrugated material of which the three sides and roof are to be constructed is $30/ft2. Design the structure so as to minimize the materials cost. Formulation. As shown in Figure 7.4, we introduce the variables x1 ⫽ depth (ft), x2 ⫽ width (ft), x3 ⫽ height (ft). The material cost will be given by Cost of roof Cost of backwall Cost of sides Thus, our objective function becomes

30x1x2 30x2x3 2(30x1x3)

316


Figure 7.4. Schematic for Example 7.3.

ƒ(x) ⫽ 30x1x2 ⫹ 30x2x3 ⫹ 60x1x3 The equality constraints consist of the volume requirement x1x2x3 ⫽ 16,000 ft3 The inequality constraints will include the perimeter limitation 2(x1 ⫹ x2) ⭐ 220 ft the width limitation x2 ⭐ 3x1 the height limitation x3 ⭐ –23 x2 and the bounds 0 ⭐ x1 ⭐ 60 ft

0 ⭐ x2 ⭐ 80 ft

In standard NLP form, the problem becomes

7.2


Minimize

317

ƒ(x) ⫽ 30x1x2 ⫹ 30x2x3 ⫹ 60x1x3

Subject to h1(x) ⫽ x1x2x3 ⫺ 16,000 ⫽ 0 g1(x) ⫽ 110 ⫺ x1 ⫺ x2 ⭓ 0 g2(x) ⫽ 3x1 ⫺ x2 ⭓ 0 g3(x) ⫽ –23 x2 ⫺ x3 ⭓ 0 0 ⭐ x1 ⭐ 60 0 ⭐ x2 ⭐ 80 0 ⭐ x3 To solve this problem using the complex method, the equality constraint must be eliminated. In this case, explicit elimination is possible by just solving h1(x) for one of the variables, say x3: 16 ⫻ 103 x1x2 Using this equation to eliminate x3, we obtain the reduced problem x3 ⫽

Minimize

ƒ(x) ⫽ 30x1x2 ⫹

48 ⫻ 104 96 ⫻ 104 ⫹ x1 x2

Subject to g1(x) ⫽ 110 ⫺ x1 ⫺ x2 ⭓ 0 g2(x) ⫽ 3x1 ⫺ x2 ⭓ 0 g3(x) ⫽ –23 x2 ⫺

16 ⫻ 103 ⭓0 x1x2

0 ⭐ x1 ⭐ 60 0 ⭐ x2 ⭐ 70 Observe that g3 becomes nonlinear in x1 and x2. Note that since bounds are already imposed on x1 and x2, it is not necessary to impose estimated bounds for use with the complex strategy. Solution. For purposes of initiating the complex method, suppose we use the point x(1) ⫽ (50, 50) as the initial feasible solution estimate. Since this is a two-dimensional problem, it will be necessary to generate three additional points. Each such point will be determined from the relations

318


x(k) 1 ⫽ 0 ⫹ r1(60)

x2(k) ⫽ 0 ⫹ r2(70)

where r1 and r2 are random numbers uniformly distributed on the interval 0 to 1. Suppose the system pseudo-random-number routine yields r1 ⫽ 0.5 and r2 ⫽ 0.9. Thus, x(2) ⫽ (30, 63) Since this point is feasible, it will be retained as the second vertex. Suppose the next pair of sampling of the random-number generator yield r1 ⫽ 0.1

r2 ⫽ 0.15

and

Then, x(3) ⫽ (6, 10.5) As shown in Figure 7.5, this point is infeasible. To render it feasible, retract it toward the centroid of the remaining two feasible vertices x(1) and x(2). Since x ⫽ –12 (x(1) ⫹ x(2)) ⫽ (40, 56.5) we have x(3) ⫽ –12 (x(3) ⫹ x) ⫽ (23, 33.5) This point is feasible and is thus retained. Finally, suppose we obtain r1 ⫽ 0.9 and r2 ⫽ 0.33; then x(4) ⫽ (54, 23.1), which is also a feasible point. Having identified four initial feasible vertex points, we next evaluate the objective function at these points and proceed with the first normal complex iteration. Index

x1

x2

1 2 3 4

50 30 23 54

50 63 33.5 23.1

ƒ(x) 103.8 87.94 72.64 87.87

⫻ ⫻ ⫻ ⫻

103 103 103 103

From the tabulation, it is evident that x(1) is the worst point and should be reflected out. The new candidate point obtained with step size ␣ ⫽ 1.3 will be

7.2


319


x(5) ⫽ x ⫹ 1.3(x ⫺ x(1)) ⫽

冉冊再冉冊冉冊冎 35.67 ⫹ 1.3 39.87

35.67 50 ⫺ 39.87 50

⫽ (17.03, 26.69)

However, at x(5), g3 is violated. The point is retracted toward the centroid, to obtain x(5) ⫽ –12 (x(5) ⫹ x) ⫽ (26.35, 33.28) a feasible point with objective function value ƒ(x(5)) ⫽ 73.37 ⫻ 103 which is lower than the value at x(1).

320


The current set of vertices consists of x(2), x(3), x(4), and x(5). The worst of these is x(2), with the function value 87.94 ⫻ 103. The centroid of the remaining points is x ⫽ (34.45, 29.96), and the new point obtained after two successive retractions toward the centroid is x(6) ⫽ (34.81, 27.28), with ƒ(x(6)) ⫽ 77.47 ⫻ 103. Observe, from Figure 7.5, that the current vertices x(3), x(4), x(5), and x(6) are all more or less flattened along constraint g3. Subsequent iterations tend to move the search closer to the optimum solution x* ⫽ (20, 40). However, as evident from the iteration summary given in Table 7.1, the vertices tend to remain flattened against the constraint, and the search slows down considerably. 7.2.3

Discussion

While the complex method does not require continuity of the problem functions, since it makes no use of function value differences, strictly speaking it does require that the feasible region be a convex set (see Appendix A for definition). This requirement arises in two places: in the calculation of the centroid [steps 1(b) and 2(b)] and in the retraction of a feasible but unsatisfactory point. In the former case we assume that the centroid, that is, the

Table 7.1

Iteration Sequence for Example 7.3

Index

x1

x2

ƒ(x), ⫻10⫺3

Vertex Replaced

7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

23.84 17.63 19.91 18.99 21.61 19.25 19.02 18.80 18.94 19.245 19.476 20.021 20.021 20.842 21.359 21.603 21.471 21.684 21.650 21.625

32.69 36.978 34.75 35.85 33.96 35.73 35.77 36.11 35.94 35.62 35.46 35.44 35.44 35.08 35.14 35.08 35.04 35.105 35.197 35.299

72.880 72.747 72.491 72.478 72.497 72.437 72.484 72.483 72.475 72.459 72.437 72.349 72.349 72.331 72.309 72.321 72.323 72.319 72.206 72.293

4 6 5 7 8 3 11 9 13 14 10 15 16 12 17 18 19 20 23 22

7.2


321

convex combination of a set of feasible points, will itself be a feasible point. As shown in Figure 7.6a, this need not be the case for a nonconvex feasible region. Similarly, with a nonconvex region, a retracted point need not be feasible even if the centroid is feasible (see Figure 7.6b). Consequently, for nonconvex regions the method could fail to converge satisfactorily. In practice, the method is widely used and has successfully solved numerous nonconvex problems; thus, the above situations must arise with low probability. As shown in Example 7.3, the complex method can slow down considerably once the trial points distribute along a constraint. In fact, in Example 7.3, with any reasonable termination criteria the algorithm might have terminated well short of the optimum. This behavior is not at all atypical, and consequently many implementations of this method will allow several terminations and subsequent restarts to occur before actually terminating the search. Thus, for the first several times the termination criterion is met, the current best point is saved, and the entire search is restarted from this point. The search is terminated only after a specified number of restarts. This device seems to improve the overall convergence rate of the algorithm. Of course, neither this nor any other heuristic construction can guarantee that the method will always converge to a local minimum. Although the method generally is slow, it will reliably yield some improvement over the supplied starting point and thus has become widely used in engineering applications of modest variable dimensions. As reviewed by Swann (ref. 6, Chap. 7), numerous authors have presented enhancements to Box’s basic algorithm. These have included more elaborate expansion/contraction rules, use of weighted centroids, modifications to accommodate infeasible centroids, simplex generation constructions that do not use random sampling, and use of quadratic interpolation constructions [4]. In general, these enhancements of the rudimentary algorithm have not resulted

Figure 7.6. Consequence of nonconvex feasible region: (a) centroid; (b) retraction.

322


in significant improvement. If a more sophisticated algorithm is appropriate or needed, then the methods of Chapter 9 or 10 are clearly to be preferred to an elaborate heuristic method.

7.3

RANDOM-SEARCH METHODS

Given that random sampling procedures have been applied to locate feasible starting points and to initiate the complex method, it is reasonable to consider the development of direct-search procedures based wholly or largely on random sampling. In fact, since the work reported by Brooks [9] in 1958, proposals employing random point generation strategies have proliferated in the engineering literature. These proposals range from simple simultaneous sampling techniques to more elaborate procedures that involve a coupling of sequential sampling with heuristic hill-climbing methods. In this section we consider briefly several examples of search methods based on random sampling. We find that while sampling techniques can be effective in locating the vicinity of the optimum, they are quite inefficient if a closer estimate of the solution is required. 7.3.1

Direct Sampling Procedures

The most elementary form of sampling procedure is one in which trial points are generated via Eq. (7.1). Each generated trial point is tested by evaluating the inequality constraints, and if it is found to be feasible, its objective value is compared to the best value currently available. If the current point yields a better value, it is retained; if not, it is rejected. The process continues until a specified number of points have been generated or the allocated CPU time has been exhausted. This procedure is referred to as simultaneous search and is known to be quite inefficient in its use of function evaluations. An analysis given by Spang [10] suggests that in order to attain 90 percent confidence ⫺ x(L) that the range of uncertainty for each variable x1 is reduced from x(U) 1 1 (U) (L) to some specified value ␧i(xi ⫺ xi )(0 ⬍ ␧i ⬍ 1), a simultaneous random search will require on the order of

写␧ N

2.3

⫺1 i

i⫽1

trial values. If N ⫽ 5 and ␧i ⫽ 0.01, this estimate indicates that the number of trial points required will be about 2.3 ⫻ 1010. A purely simultaneous search is thus clearly unsatisfactory. A somewhat more efficient approach is to divide the sampling into a series of simultaneous blocks. The best point of each block is used to initiate the

7.3


323

next block which is executed with a reduced variable sampling range. In the variant of this strategy proposed in reference 11, the user selects the number of points per block P and the number of blocks Q and proceeds as outlined below. Sampling with Interval Reduction. Given an initial feasible solution x(0), an initial range estimate z(0), where (x 0 ⫺ –12 z(0)) ⭐ x ⬍ (x(0) ⫹ –12 z(0)), and a range reduction factor ␧, 0 ⬍ ␧ ⬍ 1. For each of Q blocks, and initially with q ⫽

1, we execute the following steps: Step 1. For i ⫽ 1, . . . , N, sample the uniform distribution on (⫺ –12 , –12 ) and calculate x pi ⫽ x q⫺1 ⫹ rziq⫺1. i Step 2. If x p is infeasible and p ⬍ P, repeat step 1. If x p is feasible, save x p and ƒ(x p), and if p ⬍ P, repeat step 1. If p ⫽ P, then set x q to be the point that has the lowest ƒ(x p) over all feasible x p. Step 3. Reduce the range via zqi ⫽ (1 ⫺ ␧)zq⫺1 . i Step 4. If q ⬎ Q, terminate the search. Otherwise increment q and continue with step 1. The authors suggest a value of ␧ ⫽ 0.05 and blocks of 100 points each. The number of blocks is related to the desired reduction in the variable uncertainty. With 200 blocks, the initial range will be reduced by (1 ⫺ 0.05)200 ⯝ 3.5 ⫻ 10⫺5 at the cost of 2 ⫻ 104 trial evaluations. For simple low-dimensionality problems in which each evaluation of the set of constraints and the objective function may require only 10⫺3 s or less CPU time, the total computational burden may well be acceptable. For more realistic models—for instance, those requiring execution of a simulation—this type of sampling is prohibitively expensive. Example 7.4

Fuel Allocation in Power Plants [11]

Each of two electric power generators can be fired with either fuel oil or fuel gas or any combination of the two. The combined power that must be produced by the generators is 50 MW. The fuel gas availability is limited to 10 units/ph. It is desired to select the fuel mix for each generator so as to minimize the fuel oil utilization. From the curve fitting of operating data, the fuel requirements of generator 1 to attain a power output of x1 megawatts can be expressed as

324


w1 ⫽ 1.4609 ⫹ 0.15186x1 ⫹ 0.00145x12

(7.2)

w2 ⫽ 1.5742 ⫹ 0.1631x1 ⫹ 0.001358x21

(7.3)

where w1 and w2 are, respectively, fuel oil and fuel gas requirements in units per hour. Similarly for generator 2, to attain x2 megawatts of power, the fuel requirements are y1 ⫽ 0.8008 ⫹ 0.2031x2 ⫹ 0.000916x22

(7.4)

y2 ⫽ 0.7266 ⫹ 0.2256x2 ⫹ 0.000778x22

(7.5)

where y1 and y2 are, respectively, fuel oil and fuel gas requirements in units per hour. The operating range of generator 1 is restricted to 18 ⭐ x1 ⭐ 30

(7.6)

14 ⭐ x2 ⭐ 25

(7.7)

and that of generator 2 to

Suppose the fuels can be combined in an additive fashion. That is, for a given value of power output x*, 1 any linear combination of fuel utilization rates ␭1w1(x*) 1 ⫹ (1 ⫺ ␭1)w2(x*) 1

where 0 ⭐ ␭1 ⭐ 1, will also produce x*1 of power. A similar assumption holds for the second generator. The problem thus can be posed as follows: Determine the output rates of each generator x1 and x2 and the fuel mix fractions ␭1 and ␭2 so as to minimize the total oil consumption: ƒ ⫽ ␭1w1(x1) ⫹ ␭2 y1(x1) subject to the constraints on the availability of gas, (1 ⫺ ␭1)w2(x1) ⫹ (1 ⫺ ␭2)y2(x2) ⭐ 10

(7.8)

the total requirement, x1 ⫹ x2 ⫽ 50 and the variable bounds (7.6) and (7.7) as well as

(7.9)

7.3

0 ⭐ ␭1 ⭐ 1

and


0 ⭐ ␭2 ⭐ 1

325

(7.10)

Equality constraints (7.9) can clearly be used to solve for x2, so that x2 can be eliminated from the formulation. The resulting problem involves three variables, one nonlinear inequality (7.8), and the variable upper and lower bounds (7.10), as well as 25 ⭐ x1 ⭐ 30 which results when (7.6), (7.7), and (7.9) are combined. It is shown in reference 11 that with an initial solution estimate of x1 ⫽ 20

␭1 ⫽ –12

␭2 ⫽ –12

and the initial range estimate z1 ⫽ 20

z2 ⫽ 1.0

z3 ⫽ 1.0

a solution within 0.1 percent of the optimum objective function value is obtained in 55 blocks of 100 trial points each. To obtain a solution within 0.01 percent of the optimum, a total of 86 blocks were required, or 8600 trial points. The optimum solution obtained is ƒ ⫽ 3.05 fuel oil units/hr x1 ⫽ 30.00 ␭1 ⫽ 0.00

(x2 ⫽ 20.00) and

␭2 ⫽ 0.58

Obviously, the requirement of 5500 trial points to locate the optimum of a problem involving a cubic objective function, one cubic inequality, and three bounded variables is quite excessive. Moreover, the requirement of 3100 additional trial points to improve the solution accuracy by one additional significant figure is unsatisfactory. While the above heuristic sampling method is based on the direct reduction of the variable range with each fixed block of trials, it is possible to force convergence more effectively by modifying the distribution function of the random variable. For instance, Gaddy and co-workers [12, 13] have advocated the use of the formula. x pi ⫽ xi ⫹ zi(2r ⫺ 1)k where x is the current best solution, zi is the allowable range of variable i, and the random variable r is uniformly distributed on the interval 0 to 1.

326


When k ⫽ 1, then in effect x pi becomes a random variable uniformly distributed on the interval (xi ⫺ zi, xi ⫹ zi). When k ⫽ 3, 5, 7, . . . , the distribution of x pi becomes more and more concentrated around xi. Thus k can serve as an adaptive parameter whose value will regulate the contraction or expansion of the search region. In reference 12 an adjustment procedure is recommended in which k increased by 2 whenever a specified number of improved points is found and decreased by 2 when no improvement is found after a certain number of trials. An advantage of this approach is that the number of trials between adjustments is flexible, varying with the progress of the search. Further refinements of this type of search are reported in reference 14, in which new points are generated by sampling from a normal distribution centered at the current best point x with the variance of the distribution as the heuristically adjusted parameter. For instance, the variance is reduced by onethird after each improved point and perhaps increased by a factor of 2 after, say, 100 unsuccessful trial points. Empirical evidence given in both references 12 and 14 indicates that these methods are reasonably efficient for solving low-dimensionality problems, provided that accuracies within 1 percent of the optimum solution value are acceptable. However, the efficiency deteriorates significantly if more accurate solutions are sought. For instance, to increase the solution accuracy by one significant figure, the number of iterations must be tripled [14]. Use of such strategies in optimizing a six-variable chemical process simulation model is reported by Doering and Gaddy [13]. Typically, several hundred trial simulations were required to reach an optimum from fairly close initial solution estimates. On the whole, the body of available evidence indicates that the above simple techniques based on random sampling should not be used when problems of larger dimensions must be solved or more accurate solutions are required. 7.3.2

Combined Heuristic Procedures

To improve the performance of methods based on random sampling, various authors have proposed the incorporation of more complex adaptive or heuristic rules that serve to inject some hill-climbing moves into the overall search. These methods range from simple randomized descent schemes [11], to more complex adaptive range reduction methods [15], to rather elaborate partialenumeration strategies [16]. In this section we briefly consider only two representative strategies. The literature of the various engineering disciplines contains numerous reports of similar types of search schemes, most presented without extensive comparative testing to verify their relative effectiveness (see for example, the review given in ref. 17). The two variants discussed here were selected primarily because they illustrate a certain type of strategy and not necessarily because they are numerically superior to other similar search methods based on sampling. In the adaptive step-size random search [16], originally posed for unconstrained problems, sampling is used to generate a search direction, while the step length in that direction is determined by the previous history of successes

7.3


327

and failures. Specifically, if two successive steps lead to improvements, the step size is increased by a factor ␣s. If M successive steps do not result in improvement, the step size is reduced by the factor ␣ƒ. In outline form, and with suitable modifications to accommodate constraints, the algorithm proceeds as follows. Adaptive Step-Size Random Search. Given parameters ␣s, ␣ƒ, and M and an initial feasible point x 0 set the initial step size ␣ to 1.0 and the failure counter m to zero.

Step 1. Generate a random-direction vector d of unit length and calculate x(1) ⫽ x(0) ⫹ ␣d Step 2. If x(1) is feasible and ƒ(x(1)) ⬍ ƒ(x(0)), set y ⫽ x(0) ⫹ ␣s(x(1) ⫺ x(0)) and go to step 3. Otherwise, set m ⫽ m ⫹ 1 and go to step 4. Step 3. If y is feasible and ƒ(y) ⬍ ƒ(x(0)), then set ␣ ⫽ ␣s ␣, x ⫽ y, and go to step 5. Otherwise set x(0) ⫽ x(1) and go to step 1. Step 4. If m ⬎ M, set ␣ ⫽ ␣ƒ␣, m ⫽ 0, and go to step 5. Otherwise, continue with step 5. Step 5. Check for termination. If the termination rule is not met, go to step 1. Recommended values of the parameters of the method [16] are ␣s ⫽ 1.618, ␣ƒ ⫽ 0.618, and M ⫽ 3N, where N is the problem dimension. A numerical investigation of several random searches [18], including the above, which used only unconstrained test problems indicated that the adaptive step-size random-search procedure was quite effective in reducing the objective function during the initial search phases with problems up to 10 variables. For more precise solutions, the average linear convergence rate was rather slow. Although comparisons with the strategies of references 11, 12, and 14 are not available, the more sequential nature of the adaptive directionbased strategy would appear to make it preferable. By way of contrast, the combinatorial heuristic method [19] advanced for the solution of complex optimal mechanism-design problems takes the approach of discretizing the range of the independent variables and then carrying out a randomized search over the resulting finite grid of possible solutions. The discretized grid of solutions is constructed by simply dividing the expected range of each independent variable into a suitable number of fixed variable values. The search over this fixed grid of variable values is based on a semirandomized one-variable-at-a-time strategy. The strategy is given here in outline form. Combinatorial Heuristic Method

Step 1. Generate a random feasible starting point x0 and set Fmin ⫽ ƒ(x 0). For each variable i, i ⫽ 1, 2, . . . , N, execute the following loop of steps.

328


Step 2. Optimize the ith variable, holding the others fixed. (a) Conduct a randomized selection of values of the variable i to determine q additional feasible values that yield a better objective function value than the current base point. If none can be found, reinitiate step 2 with variable i ⫹ 1. (b) Identify the best of the q feasible solutions and set its objective function value to T min. (c) Conduct a look-ahead search: (i) For each of the q feasible solutions identified in step 2(a), conduct a randomized selection of q values of the (i ⫹ 1)st variable to determine the feasible value of that variable that yields an objective function improvement over Tmin. (ii) Select the best of these q feasible points. The value of the ith variable of that point is fixed as being optimum. (d) If i ⫽ N, go to step 3. Otherwise, go to step 2(a) and consider the (i ⫹ 1)st variable. Step 3. Perform a randomized search to identify an improved value of the Nth variable, holding all other variables at their current base-point values. The result becomes the new base point, and its function value is the new Fmin. Step 4. Check for termination. If the termination criteria are not met, reinitiate step 2 with i ⫽ 1. Several termination criteria could be employed in step 4, including the normal one of insufficient improvement in the objective function value. The authors of the method suggest that q, the number of look-ahead points, should be between 3 and 5. The reader should note that once the random sampling to determine the starting point is completed, the remaining sampling involves only the selection of discrete values of a given variable. The sampling in this case only amounts to an unbiased way of choosing which discrete value to test next and is carried out so as to avoid an exhaustive search over all discrete values of the variable in question. The normal technique for carrying out this type of randomized selection is to assign an equal subinterval of the unit interval to each discrete value of the variable, to sample the uniform distribution on the interval (0, 1), and then to select the discrete variable value based on the subinterval within which the sampled random variable value is located. Further enhancements of the combinatorial heuristic method reported by Datseris [20] include variable reordering, more elaborate look-ahead procedures, and variable reduction strategies. The proponents of the combinatorial heuristic method have reported on the application of this technique to several mechanism-design problems, including

7.3


329

the design of a variable-speed drive, a recording mechanism, a twin crank drive, and a two-mesh gear train [20, 21]. In these cases the method yield improved ‘‘optimal’’ solutions over designs previously reported in the literature. However, although some limited evidence is available that this technique is more efficient than a direct simultaneous random search, extensive comparative results on the basis of which an evaluation could be made are not available.

7.3.3

Discussion

The available fragmentary evidence indicates that random-sampling-based search methods ought to be used either with low-dimensionality problems or as devices for generating good starting points for more sophisticated optimization methods. It is true, however, that any sampling strategy that covers most of the feasible region inherently offers enhanced possibilities of locating the vicinity of the global optimum. In fact, research is under way in the unconstrained case to exploit this feature in solving multimodal optimization problems [22–24]. However, this work also indicates that sampling methods should be used only to identify the approximate neighborhood of the minimum and that theoretically based methods should be employed to refine this estimate. Of the random-search methods described in this section, those using some adaptive heuristics are likely to be more effective because they minimize the use of simultaneous samples. In general, a sequential sampling procedure will be more efficient than a simultaneous sampling procedure, because it can make use of the results of previous test points. Moreover, sampling procedure such as that of Luus and Jaakola [11], which successively force a reduction of the feasible region, must be used with some caution to ensure that the reduction heuristic does not eliminate the optimal solution itself. It is quite possible that many or most of the trial points employed per block can be infeasible, in which case the base point for the next block may be selected from very few feasible points and thus may direct the search toward suboptimal regions. With reduction-based strategies, the number of feasible points found per block must be monitored and the total number of trials per block increased if the count of the former is low. Although systematic evidence is not available, the general impression gained from the literature is that random-search methods can be quite effective for severely nonlinear problems that involve multiple local minima but are of low dimensionality. However, for problems of more dimensions with many inequality constraints, the methods of the next three chapters generally are much more efficient.

330

7.4


SUMMARY

This chapter has focused on techniques for locating constrained optima that rely only upon the use of objective and constraint function values. Because of the difficulty of directly satisfying nonlinear equality constraints, this class of methods requires that the problem be reformulated so that equality constraints are explicitly or implicitly eliminated. Moreover, these methods normally require that a feasible starting point be available. We found that random sampling was one possible way of generating such points. The methods themselves can be divided into adaptations of unconstrained direct-search methods and strategies based on random-search procedures. We found that the algorithms that employ searches along a fixed set of directions invariably encounter situations that lead to premature termination. Thus, such methods necessarily require modifications that redirect the set of directions to take the constraint normals into account. The extension of the simplex search to accommodate constraints, on the other hand, requires only minor revisions, but, theoretically, these extensions could fail if the feasible region is not convex. Random-search procedures range from direct sampling and rejection methods to more complex strategies that employ sampling together with various heuristic or adaptive devices. We found that sequential sampling is always preferable to simultaneous sampling and that with the sequential procedures some mechanism of range reduction is quite expedient. In general, simple random-sampling-based techniques are best applied to low-dimensionality problems and in applications in which highly accurate solutions are not required. The chapter concluded with a brief examination of combined sampling and heuristic procedures. The adaptive step-size random search uses sampling to define a search direction but heuristics to adjust the step size. The combinatorial heuristic method is based on a discretization of the independent variable ranges and a randomized search of the resulting grid of trial points. Available fragmentary evidence suggests that such more elaborate sampling procedures can be quite effective for low-dimensionality, severely nonlinear problems. Random-search-based methods are probably best used to generate good feasible starting points and to identify the vicinity of global solutions for problems with multiple local optima.

REFERENCES 1. Wait, R. The Numerical Solution of Algebraic Equations, Wiley, New York, 1979. 2. Westerberg, A. W., H. P. Hutchinson, R. L. Motard, and P. Winter, Process Flowsheeting, Cambridge University Press, Cambridge, England, 1979. 3. Sargent, R. W. H., ‘‘A Review of Methods for Solving Nonlinear Algebraic Equations,’’ in Foundations of Computer Aided Process Design, Vol. 1 (R. S. H. Mah and W. D. Seider, Eds.) Engineering Foundation, New York, 1981, pp. 27–76.

REFERENCES

331

4. Box, M. J., D. Davies, and W. H. Swann, Nonlinear Optimization Techniques, ICI Monograph No. 5, Oliver & Boyd, London, 1969. 5. Klingman, W. R., and D. M. Himmelblau, ‘‘Nonlinear Programming with the Aid of a Multiple Gradient Summation Technique,’’ JACM, 11, 400–415 (1964). 6. Gill, P. E., and W. Murray (Eds.), Numerical Methods for Constrained Optimization, Academic, London, 1974. 7. Box, M. J., ‘‘A New Method of Constrained Optimization and a Comparison with Other Methods,’’ Computer J., 8, 42–52 (1965). 8. Biles, W. E., ‘‘Optimization of Multiple Response Simulation Models,’’ University of Notre Dame, Final Report, ONR-Contract N00014-76-C-1021, Oct. 31, 1978. 9. Brooks, S. H., ‘‘A Discussion of Random Methods for Seeking Maxima,’’ Oper. Res., 6, 244–253 (1958). 10. Spang, H. A., ‘‘A Review of Minimization Techniques for Nonlinear Function,’’ SIAM Rev., 4, 343–365 (1962). 11. Luus, R., and T. H. I. Jaakola, ‘‘Optimization by Direct Search and Systematic Reduction of the Size of Search Region,’’ AIChE J., 19, 760–766 (1973). 12. Heuckroth, M. W., J. L. Gaddy, and L. D. Gaines, ‘‘An Examination of the Adaptive Random Search Technique,’’ AIChE J., 22, 744–750 (1976). 13. Doering, F. J., and J. L. Gaddy, ‘‘Optimization of the Sulfuric Acid Process with a Flowsheet Simulator,’’ Computers Chem. Eng., 4, 113–122 (1980). 14. Goulcher, R., and J. J. Casares Long, ‘‘The Solution of Steady State Chemical Engineering Optimization Problems Using a Random Search Algorithm,’’ Computers Chem. Eng., 2, 33–36 (1978). 15. Martin, D. L., and J. L. Gaddy, ‘‘Process Optimization with the Adaptive Randomly Directed Search,’’ in Selected Topics in Computer Aided Process Design and Analysis, CEP Symp. Ser., Vol. 214 (R. S. H. Mah and G. V. Reklaitis, Eds.), American Institute of Chemical Engineers, New York, 1982. 16. Schumer, M. A., and K. Steiglitz, ‘‘Adaptive Step Size Random Search,’’ IEEE Trans., AC-13, 270–276 (1968). 17. Jarvis, R. A., ‘‘Optimization Strategies in Adaptive Control: A Selective Survey,’’ IEEE Trans. Syst. Man and Cybern., 5, 83–94 (1975). 18. Schrack, G., and N. Borowski, ‘‘An Experimental Comparison of Three Random Searches,’’ in Numerical Methods for Nonlinear Optimization (F. A. Lootsma, Ed.), Academic, New York, 1972. 19. Lee, T. W., and F. Freudenstein, ‘‘Heuristic Combinatorial Optimization in the Kinematic Design of Mechanisms: Part 1: Theory,’’ J. Engl. Ind. Trans. ASME, 1277–1280 (1976). 20. Datseris, P., ‘‘Principles of a Heuristic Optimization Technique as an Aid in Design—An Overview,’’ ASME Tech. Paper, 80-DET-44 (Sept. 1980). 21. Lee, T. W., and R. Freudenstein, ‘‘Heuristic Combinatorial Optimization in the Kinematic Design of Mechanisms, Part 2: Applications,’’ J. Engl. Ind. Trans. ASME, 1281–1284 (1976). 22. Mockus, J., ‘‘On Baysian Methods of Optimization,’’ in Towards Global Optimization (L. C. W. Dixon and G. P. Szego, Eds.), North-Holland, Amsterdam, 1975. 23. Zilinskas, A., ‘‘On Statistical Models for Multi-Modal Optimization,’’ Math. Operationsforsh. Stat. Ser. Stat., 9, 255–266 (1978).

332


24. Devroye, L. P., ‘‘Progressive Global Random Search of Continuous Functions,’’ Math Program., 15, 330–342 (1978).

PROBLEMS

7.1. Why must equality constraints be eliminated before applying directsearch optimization methods? 7.2. Why might a penalty function construction be used to help generate a feasible starting point for an inequality-constrained problem? 7.3. What is the reason for using more than N ⫹ 1 points in the complex search method? 7.4. Why is convexity necessary to guarantee the successful operation of the complex method? 7.5. Why is restarting used in the context of the complex method? 7.6. In random sampling with interval reduction in blocks, what are the competing factors that must be considered in the choice of block size? 7.7. In interval reduction sampling procedures, is it possible to discard the minimum solution during the reduction process? If so, how can this occur, and how can it be avoided? 7.8. Which is the adaptive component of the adaptive step-size randomsearch procedure? 7.9. Show that the constraint g1(x) ⫽ x21 ⫺ x2 ⭐ 0 is convex. Suppose the constraint is rewritten as an equality by adding a slack variable. Is the constraint set g1(x) ⫽ x12 ⫺ x2 ⫹ x3 ⫽ 0 g2(x) ⫽ x3 ⭓ 0 convex? 7.10. (a) Is the set defined by the constraints g1(x) ⫽ x1 ⫹ x2 ⫺ 1 ⫽ 0 g2(x) ⫽ 2 ⫺ x21 ⫺ x22 ⭓ 0 convex?

PROBLEMS

333

(b) Is the set defined by the constraints g1(x) ⫽ x1 ⫺ 1 ⭐ 0 g2(x) ⫽ x1 ⫹ x22 ⫺ 2 ⭓ 0 convex? 7.11. Is the feasible region x1 ⫹ x2 ⫹ x3 ⫽ 1 x21 ⫹ x22 ⭐ 1

x3 ⫹ x22 ⭓ 1

a convex region? Prove your answer. 7.12. Given the nonlinear program Minimize

ƒ(x) ⫽ x21 ⫹ x22 ⫹ x23

Subject to g1(x) ⫽ 1 ⫺ x2⫺1x3 ⭓ 0 g2(x) ⫽ x1 ⫺ x3 ⭓ 0 h1(x) ⫽ x1 ⫺ x22 ⫹ x2x3 ⫺ 4 ⫽ 0 0 ⭐ x1 ⭐ 5 0 ⭐ x2 ⭐ 3 0 ⭐ x3 ⭐ 3 What transformations are necessary in order to use the complex method? Give the final transformed form. 7.13. Give the problem Minimize

ƒ(x) ⫽ ⫺x1x22x33 /81

Subject to h1(x) ⫽ x13 ⫹ x22 ⫹ x3 ⫺ 13 ⫽ 0 h2(x) ⫽ x22x3⫺1 / 2 ⫺ 1 ⫽ 0 All xi ⭓ 0 (a) Reformulate the problem for solution using the complex method. Is the resulting feasible region convex?

334


(b) Estimate upper and lower bounds on the independent variables for use in the complex method. (c) Find a starting feasible solution. 7.14. In solving the problem Minimize

3(x2 ⫺ 4)2 ⫺ 2x1

Subject to g1(x) ⫽ 10 ⫺ x12 ⫺ x22 ⭓ 0 g2(x) ⫽ 9 ⫺ x21 ⫺ (x2 ⫺ 4)2 ⭓ 0 0 ⭐ (x1, x2) ⭐ 4 the following points are generated by the complex method:

1 2 3 4

x1

x2

ƒ(x)

2.0 0.5 0.6 2.3

2.0 3.0 1.1 2.1

8 2 24.03 6.23

Assuming ␣ ⫽ 1.3, carry out two iterations of the complex method. Construct a plot of the results. 7.15. Given the problem Minimize

ƒ(x) ⫽ 3x21 ⫺ 2x2

Subject to g1(x) ⫽ 2x1 ⫹ x2 ⭓ 4 g2(x) ⫽ x21 ⫹ x22 ⭐ 40 (a) Estimate upper and lower bounds on the variables and obtain a starting feasible solution. (b) Write a program to carry out sampling with interval reduction and use it to solve the problem. Use ␧ ⫽ 0.05 and 100 blocks of 50 points each. 7.16. Solve the problem Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ (x1 ⫺ x2)2 ⫹ (x2 ⫺ x3)2

Subject to h1(x) ⫽ x1(1 ⫹ x22) ⫹ x42 ⫺ 4 ⫺ 3兹2 ⫽ 0 All xi, 0 ⭐ xi ⭐ 3

PROBLEMS

335

using a computer implementation of the complex method. [ƒ* ⫽ 0.032567; see Miele et al., J. Optim. Theory Appl., 10, 1 (1972).] 7.17. Solve the problem Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ (x1 ⫺ x2)2 ⫹ (x3 ⫺ 1)2 ⫹ (x4 ⫺ 1)4 ⫹ (x5 ⫺ 1)6

Subject to h1(x) ⫽ x12x4 ⫹ sin(x4 ⫺ x5) ⫺ 2兹2 ⫽ o h2(x) ⫽ x2 ⫹ x34x42 ⫺ 8 ⫺ 兹2 ⫽ 0 using an adaptive random-search procedure. [ƒ* ⫽ 0.24149; see Miele et al., J. Optim. Theory Appl., 10, 1 (1972).] 7.18. The minimum-weight design of the corrugated bulkheads for a tanker requires the solution of the following optimization problem (Kvalie et al., Tech. Rep., Norwegian Technical Institute, Trondheim, Norway, 1966). Minimize

ƒ(x) ⫽

5.885x4(x1 ⫹ x3) x1 ⫹ (x23 ⫺ x22)1 / 2

Subject to g1(x ⫽ x2x4(0.4x1 ⫹ –16 x3) ⫺8.94[x1 ⫹ (x32 ⫺ x22)1 / 2] ⭓ 0 1 g2(x) ⫽ x22x4(0.2x1 ⫹ –– 12 x3)

⫺2.2{8.94[x1 ⫹ (x32 ⫺ x22)1 / 2]}4 / 3 ⭓ 0

g3(x) ⫽ x4 ⫺ 0.0156x1 ⫺ 0.15 ⭓ 0 g4(x) ⫽ x4 ⫺ 0.0156x3 ⫺ 0.15 ⭓ 0 g5(x) ⫽ x4 ⫺ 1.05 ⭓ 0 g6(x) ⫽ x3 ⫺ x2 ⭓ 0 The variables x correspond to the following corrugation dimension: x1 width, x2 depth, x3 length, and x4 plate thickness. (a) Obtain a starting feasible solution by sampling using the bounds 0 ⭐ x1 ⭐ 100, 0 ⭐ x2 ⭐ 100, 0 ⭐ x3 ⭐ 100, and 0 ⭐ x4 ⭐ 5. (b) Solve the problem using a random-search method (ƒ* ⫽ 6.84241).

8 LINEARIZATION METHODS FOR CONSTRAINED PROBLEMS In the previous chapters we considered in some detail the two problem classes that optimization theory has treated most thoroughly: the unconstrained problem and the completely linear constrained problem. We found that for these two special cases very efficient algorithms can be devised that have a high degree of reliability and for which a rather complete theoretical analysis can be presented. In view of these successes, it should not be surprising that most approaches to the general constrained nonlinear problem have involved devices by means of which the general problem can be solved by exploiting the techniques developed for these two special cases. As discussed in Chapter 6, one successful approach to solving the general problem has been to reformulate it as a parameterized unconstrained problem. In this way, the constrained problem can be solved as a sequence of unconstrained problems using well-developed existing algorithms. In this and the next chapter we consider the use of a second device, linearization, by means of which the general problem can be converted to a constrained but linear problem. The use of linearization approximations allows us to solve the general problem by employing linear programming (LP) methodology either by solving one or more linear programs or else by employing elements of the linear simplex methodology in a recursive fashion. The methods discussed in this section employ the fact that a general nonlinear function ƒ(x) can be approximated in the vicinity of a point x 0 by Taylor’s expansion, ƒ(x) ⫽ ƒ(x 0) ⫹ ⵜƒ(x 0)(x ⫺ x 0) ⫹ O(㛳x ⫺ x 0㛳)2 Almost without exception, the higher order terms O(㛳x ⫺ x 0㛳2) are ignored, and hence the function ƒ(x) is approximated by the linearization of ƒ at x 0 denoted by 336


8.1

DIRECT USE OF SUCCESSIVE LINEAR PROGRAMS

337

˜ ƒ(x; x 0) ⫽ ƒ(x 0) ⫹ ⵜƒ(x 0)(x ⫺ x 0)

We will call the point x 0 the linearization point. It should be recognized that the linearization is in most instances a very gross approximation, which must be used with great caution. Nonetheless, it is an approximation that is very commonly used not only in optimization but in virtually all of engineering. The representative optimization techniques discussed in this chapter differ primarily in the manner and frequency with which the linearizations are updated and in the devices that are employed to account for the errors introduced by employing such approximations. 8.1


The simplest and most direct use of the linearization construction is to replace the general nonlinear problem with a complete linearization of all problem functions at some selected estimate of the solution. Since in this manner all problem functions are replaced by linear approximations, the resulting problem takes the form of a linear program and can be solved as such. The LP solution obtained will no doubt only be an approximation to the solution of the original problem; however, if adequate precautions are taken, it ought to be an improvement over the linearization point. We first consider the case of the linearly constrained problem and discuss the supplementary calculations required to make direct linearization effective. Then we consider the general nonlinear programming (NLP) problem. 8.1.1

Linearly Constrained Case

The linearly constrained NLP problem is that of Minimize

ƒ(x)

Subject to Ax ⭐ b x⭓0 This differs from the standard LP problem only in the presence of a nonlinear objective function ƒ(x). As in the LP case, the feasible region remains a polyhedron, that is, a geometric entity formed by the intersection of a set of planes. Because ƒ(x) is nonlinear, the optimal solution no longer needs to be confined to vertices or corner points of the feasible region but can instead lie anywhere within the region. Moreover, if ƒ(x) is nonconvex, the linearly constrained NLP problem may have multiple local minima. Given these problem features, let us consider the consequences of formulating and solving the linearization of the linearly constrained nonlinear program constructed at some feasible point x 0. Clearly, the problem

338

LINEARIZATION METHODS FOR CONSTRAINED PROBLEMS

Minimize

˜ ƒ(x; x 0)

Subject to Ax ⭐ b x⭓0 is an LP problem and as such, assuming the feasible region is bounded, will possess an optimal solution that is a feasible corner point. The key question is, how is x˜ *, the solution of the approximate problem, related to x*, the solution of the original problem? First of all, ignoring for the moment the case of an infinite number of LP solutions, we have reduced the possible multiple local minima to a single approximate solution. To which of the local minima of the NLP the point x˜ * will be closest is totally determined by our choice of linearization point. Hence, in the nonconvex case, we cannot be assured of finding the global minimum. However, even if ƒ(x) is convex, we have no real assurance that x˜ * will in fact even be close to x*. While x˜ * will always be a corner point, x* can lie anywhere; it may even be an interior point. Quite obviously, even if ƒ(x) is convex, some further adjustment of x˜ * will be required to ensure that we can draw nearer to x*. Note that both x 0 and x˜ * are feasible points and that by virtue of the minimization of ƒ˜ (x; x 0) it must be true that ˜ 0; x 0) ⬎ ƒ(˜ ˜ x*; x 0) ƒ(x

Consequently, if we substitute, using the linearization formula, it follows that ƒ(x 0) ⬎ ƒ(x 0) ⫹ ⵜƒ(x 0)(˜x* ⫺ x 0) or ⵜƒ(x 0)(˜x* ⫺ x 0) ⬍ 0

Evidently the vector x˜ * ⫺ x 0 is a descent direction. Recall for our discussion of the unconstrained case that while a descent direction is a desirable choice for a search direction, it can lead to an improved point only if it is coupled with a step adjustment procedure or, better yet, a line search. Note that a line search from x 0 in the direction x˜ * ⫺ x 0 will eventually lead to the point x˜ *. Since x˜ * is a corner point of the feasible region, and since x 0 is feasible, all points on the line between them will be feasible (since the feasible region is convex). Moreover, since x˜ * is a corner point, points on the line beyond x˜ * will lie outside the feasible region. The line search can thus be confined to the bounded line segment x ⫽ x 0 ⫹ ␣(˜x* ⫺ x 0)

0⭐␣⭐1

8.1


339

Thus while the LP solution does not directly yield an improved estimate of the optimum, it does accomplish two important functions: It defines a search direction and it defines the point at which the feasible region will be intersected. As a consequence of solving the problem min ƒ(x 0 ⫹ ␣(˜x* ⫺ x 0))

0⭐␣⭐1

a feasible point x1 will be found with the property ƒ(x1) ⬍ ƒ(x 0) Since ⵜƒ(x1) will in general not be equal to the zero vector, the improved point x1 will serve as a linearization point for the next approximating linear program. The sequence of alternate LP solutions followed by line searching must be continued until successive line search optima x (t) differ by less than some selected convergence tolerance. The composite algorithm, attributed to Frank and Wolfe [1], is summarized as follows. Frank–Wolfe Algorithm Given x 0, line search, and overall convergence tolerances ␧ ⬎ 0 and ␦ ⬎ 0. Step 1. Calculate ⵜƒ(x (t)). If 㛳ⵜƒ(x (t)㛳 ⭐ ␧, stop. Otherwise, go to step 2. Step 2. Solve the LP subproblem Minimize

ⵜƒ(x (t))y

Subject to Ay ⭐ b y⭓0 Let y (t) be the optimal solution to the LP problem. Step 3. Find ␣ (t) which solves the problem min ƒ(x (t) ⫹ ␣(y (t) ⫺ x (t)))

0⭐␣⭐1

Step 4. Calculate x (t⫹1) ⫽ x (t) ⫹ ␣ (t)(y (t) ⫺ x (t)) Step 5. Convergence check. If 㛳x (t⫹1) ⫺ x (t)㛳 ⬍ ␦㛳x (t⫹1)㛳

340


and if 㛳ƒ(x (t⫹1)) ⫺ ƒ(x (t))㛳 ⭐ ␧㛳ƒ(x (t⫹1))㛳

then terminate. Otherwise, go to step 1. Note that in the statement of the LP subproblem we have deleted the constant terms of the linear objective function ƒ˜ (x; x (t)) and retained only the variable portion. Example 8.1

Multistage Compressor Optimization

A gas flowing at a rate of N moles per hour at 1 atm pressure is to be compressed to 64 atm using the three-stage compressor shown in Figure 8.1. Assume that compression occurs reversibly and adiabatically and that after each stage of compression the gas is cooled back to its initial temperature T. Choose the interstage pressures so as to minimize the energy consumption. Solution. For reversible adiabatic compression with cooling to inlet temperature T, the work is given by

冉冊冉冊

W ⫽ NRT

k

k⫺1

Pout Pin

冉冊

(k⫺1) / k

⫺ NRT

k k⫺1

where k ⫽ Cp /Cv, ratio of the gas heat capacities R ⫽ ideal gas constant For three-stage compression, the total work is given by

冉冊再冉冊冉冊冉冊冎

Wtotal ⫽ NRT

k

k⫺1

x1 1

␣

⫹

x2 x1

␣

⫹

64 x2

␣

⫺3

Figure 8.1. Multistage compressor schematic, Example 8.1.

8.1


341

where ␣ ⫽ k ⫺ 1/k x1 ⫽ outlet pressure from first stage x2 ⫽ outlet pressure from second stage If for the gas in question ␣ ⫽ –14 , then for fixed T and N the optimal pressures x1 and x2 will be obtained as the solution of the problem Minimize

ƒ(x) ⫽ x 11 / 4 ⫹

Subject to x1 ⭓ 1

冉冊冉冊 x2 x1

x2 ⭓ x1

1/4

⫹

64 x2

1/4

64 ⭓ x2

where the constraints are imposed to ensure that the gas pressures selected will be monotonically increasing from inlet to outlet. This is a linearly constrained NLP problem and can be solved using the Frank–Wolfe algorithm. Suppose as our initial estimate we use x 01 ⫽ 2 and x 02 ⫽ 10, a feasible solution. The gradient of ƒ(x) at x 0 is ⭸ƒ /4 ⫽ 0.25x⫺3 (1 ⫺ x 21 / 4x1⫺1 / 2) ⫽ ⫺3.83 ⫻ 10⫺2 1 ⭸x1 ⭸ƒ / 4 ⫺1 / 4 /2 ⫽ 0.25x⫺3 [x1 ⫺ (64)1 / 4x⫺1 ] ⫽ ⫺2.38 ⫻ 10⫺3 2 2 ⭸x2

Assuming that these values are not sufficiently close to zero, we continue with step 2: Solve the LP problem Minimize

⫺3.83 ⫻ 10⫺2y1 ⫺ 2.38 ⫻ 10⫺3y2

Subject to y1 ⭓ 1

y2 ⭓ y1

64 ⭓ y2

The minimum will clearly occur when y1 and y2 are as large as possible, namely y 01 ⫽ y 02 ⫽ 64. Step 3. Search the line

冉冊冉冊冉冉冊冉冊冊 x1 2 ⫽ ⫹␣ x2 10

64 2 ⫺ 64 10

for 0 ⭐ ␣ ⭐ 1. Using, say, an interpolation method, we find that the optimum occurs at ␣ ⫽ 0.02732. Step 4. The new point is

冉冊冉冊

冉冊冉冊

x (1) 2 62 3.694 1 ⫽ ⫹ 0.02732 ⫽ x (1) 10 54 11.475 2

342


As shown in Figure 8.2, the point x (1) is an interior point of the feasible region. Since the search clearly has not yet converged, the iterations resume with step 1. Step 1. ⵜƒ(x (1)) ⫽ (3.97 ⫻ 10⫺3, ⫺4.56 ⫻ 10⫺3)

Step 2. The LP subproblem becomes Minimize

⫹3.97 ⫻ 10⫺3y1 ⫺ 4.56 ⫻ 10⫺3y2

Subject to y1 ⭓ 1

y2 ⭓ y1

64 ⭓ y2

The minimum is attained at the corner point at which y1 is as small and y2 is as large as possible; thus, y (1) 1 ⫽ 1

and

y 2(1) ⫽ 64

Figure 8.2. Feasible region, Example 8.1.

8.1


343

Step 3. Search the line

冉冊冉冊冉冉冊冉冊冊 x1 3.694 ⫽ ⫹␣ x2 11.475

1 3.694 ⫺ 64 11.475

for 0 ⭐ ␣ ⭐ 1 such that ƒ(x(␣)) is minimized. The optimum occurs at ␣ ⫽ 0.06225. Step 4. The new point is

冉冊冉冊

冉

冊冉冊

x (2) 3.694 ⫺2.694 3.526 1 ⫽ ⫹ 0.06225 ⫽ x (2) 11.475 52.525 14.745 2

From Figure 8.2, this point is also an interior point. Moreover, it is still sufficiently distant from the optimum. The process continues with the points x (3) ⫽ (3.924, 15.069)

x (4) ⫽ (3.886, 15.710)

and so on, until suitably tight convergence tolerances are met. From Figure 8.2 it is evident that with an interior optimum the Frank– Wolfe algorithm advances very similarly to the unconstrained gradient method. In fact, the zigzagging along a limited set of directions is typical of the ordinary gradient method. If both the starting point and the solution are boundary points, then the search will eventually reduce to a sequence of oneparameter searches on the constraints. Example 8.2

Suppose that the compressor system of the previous example is constrained to satisfy the additional condition x2 ⭐ 14 atm Obtain the new optimum using as initial estimate x 0 ⫽ (14, 14). Solution. The given initial point is a corner point of the modified feasible region, as shown in Figure 8.3. At x 0, ⵜƒ(x 0) ⫽ (1.67 ⫻ 10⫺2, ⫺8.25 ⫻ 10⫺3)

The corresponding LP subproblem is

344



Minimize

1.67 ⫻ 10⫺2y1 ⫺ 8.25 ⫻ 10⫺3y2

Subject to y1 ⭓ 1

y2 ⭓ y1

14 ⭓ y2

The solution is obtained when y1 is as small and y2 is as large as possible. Thus, y 01 ⫽ 1

y 02 ⫽ 14

and

Step 3, the line search, involves finding ␣, 0 ⭐ ␣ ⭐ 1, such that

冉冊冉冊冉冉冊冉冊冊 x1 14 ⫽ ⫹␣ x2 14

1 14 ⫺ 14 14

and ƒ(x(␣)) is minimized. This amounts to a search on the constraint y2 ⫽ 14. The solution is ␣ ⫽ 0.7891

and the new point is

8.1


冉冊冉

345

冊

x (1) 3.7417 1 ⫽ x (1) 14.0 2

At this point, ⵜƒ(x (1)) ⫽ (0, ⫺1.275 ⫻ 10⫺3). This is the optimal solution and, as shown in Figure 8.3, it lies in the middle of the constraint y2 ⫽ 14. Note that if the additional constraint x1 ⭓ 5 were imposed (dashed line in Figure 8.3), then the optimal solution would be the corner point (5, 14). From the example, it is clear that even for boundary points the Frank– Wolfe algorithm amounts to a gradient search projected onto the constraint surfaces. Convergence to a Kuhn–Tucker point from any feasible starting point can readily be proved (see Zangwill [2] or Wolfe [3]) under reasonable assumptions: continuously differentiable ƒ(x), bounded feasible region, and accurate LP and line search subproblem solutions. If ƒ(x) is convex, clearly the solution will be the global minimum. It is awkward to analyze the rate of convergence because of the involvement of the vertex points in the definition of the descent direction. However, it is sufficiently clear that there are no grounds to expect a better rate than that of the ordinary gradient method. Finally, it is interesting to note that with the Frank–Wolfe algorithms (in fact with any algorithm that uses a subproblem with a linearized objective function), if the objective function is convex, then after each subproblem solution it is possible to obtain a very convenient estimate of the remaining improvement yet to be achieved in the value of the objective function. Recall that for a convex function linearized at any point x (t) ˜ ƒ(x) ⭓ ƒ(x (t)) ⫹ ⵜƒ(x (t))(x ⫺ x (t)) ⬅ ƒ(x; x (t))

for all points x. But this inequality implies that over all feasible points ˜ ˜ (t); x (t)) min ƒ(x) ⭓ min ƒ(x; x (t)) ⬅ ƒ(y

Hence, the value of ƒ˜ (y (t); x (t)) is a lower bound estimate of the value of ƒ(x) at the minimum point. On the other hand, ƒ(y (t)), or better yet ƒ(x (t⫹1)), is an upper bound estimate of the minimum value of ƒ(x). Thus, after each cycle of the algorithm, the difference ˜ (t); x (t)) ƒ(x (t⫹1)) ⫺ ƒ(y

gives a conservative estimate of the attainable improvement in the objective function value. This difference can thus be used as a good termination criterion, supplementing those defined in step 5 of the Frank–Wolfe algorithm. With this basic introduction to the use of direct linearization in solving the special case of linearly constrained NLP problems, we continue with a discussion of the general case.

346


8.1.2

General Nonlinear Programming Case

Consider the general NLP problem Minimize

ƒ(x)

Subject to

gj (x) ⭓ 0

j ⫽ 1, . . . , J

hk(x) ⫽ 0

k ⫽ 1, . . . , K

x i(U) ⭓ xi ⭓ x i(L)

i ⫽ 1, . . . , N

Given an estimate x (t), the direct linear approximation problem that can be constructed at x (t) is Minimize

ƒ(x (t)) ⫹ ⵜƒ(x (t))(x ⫺ x (t))

Subject to

gj (x (t)) ⫹ ⵜgj (x (t))(x ⫺ x (t)) ⭓ 0

j ⫽ 1, . . . , J

hk(x (t)) ⫹ ⵜhk(x (t))(x ⫺ x (t)) ⫽ 0 x

(U) i

⭓ xi ⭓ x

k ⫽ 1, . . . , K (L) i

i ⫽ 1, . . . , N

Clearly this is an LP problem and as such can be solved using LP methods to give a new point x (t⫹1). Note, however, that even if x (t) is a feasible point of the original nonlinear problem, there is no assurance that x (t⫹1) will be feasible. In fact, if the constraint functions are nonlinear, the point x (t⫹1) will almost assuredly be infeasible. But if x (t⫹1) is infeasible, then the attainment of an improvement in the value of the approximate objective function, that is, ˜ (t⫹1); x (t)) ⬍ ƒ(x ˜ (t); x (t)) ƒ(x

is no assurance that an improved estimate of the true optimum solution has been attained. In general, if {x (t)} is a series of points each of which is the solution of the LP problem obtained by using the previous LP solution as the linearization point, then in order to attain convergence to the true optimum solution it is sufficient that at each point x (t) an improvement be made in both the objective function value and the constraint infeasibility. As shown in the next example, these desirable conditions can occur in direct applications of the successive linear approximation strategy without further embellishments or safeguards.

8.1


347

Example 8.3


ƒ(x) ⫽ x 21 ⫹ x 22

Subject to g1(x) ⫽ x 12 ⫺ x2 ⭓ 0 h1(x) ⫽ 2 ⫺ x1 ⫺ x 22 ⫽ 0 –52 ⭓ x1 ⭓ –12

3 ⭓ x2 ⭓ 0

As shown in Figure 8.4a, the points satisfying the problem constraints consist of all points on the curve h1(x) ⫽ 0 between the point (2, 0) determined by the linear bound x2 ⭓ 0 and the point (1, 1) determined by constraint g1(x) ⭓ 0. Suppose we construct the linearized approximation to this problem at the point x 0 ⫽ (2, 1). The result is Minimize

˜ ƒ(x; x 0) ⫽ 5 ⫹ 4(x1 ⫹ 2) ⫹ 2(x2 ⫺ 1)

Subject to g˜ 1(x; x 0) ⫽ 3 ⫹ 4(x1 ⫺ 2) ⫺ (x2 ⫺ 1) ⭓ 0 h˜ 1(x; x 0) ⫽ ⫺1 ⫺(x1 ⫺ 2) ⫺ 2(x2 ⫺ 1) ⫽ 0 –25 ⭓ x1 ⭓ –21

3 ⭓ x2 ⭓ 0

As shown in Figure 8.4b, all points satisfying the linearized constraints will lie on the line 3 ⫺ x1 ⫺ 2x2 ⫽ 0 between the point (2.5, 0.25) determined

Figure 8.4. Feasible region, Example 8.3: (a) constraints; (b) linearized constraints.

348


11 8 by the bound –52 ⭓ x1 and the point (–– 9 , – 9 ) determined by the intersection with 0 g˜ 1(x; x ). While the original objective function contours are circles with center at the origin, the contours of the linearized objective functions are straight lines of slope ⫺2. Clearly, the solution to the linear approximating problem 11 8 is x (1) ⫽ (–– 9 , – 9 ). At this point,

g1(x (1)) ⫽ 0.6049 ⬎ 0

h1(x (1)) ⫽ ⫺0.0123 ⫽ 0

that is, the equality constraint is violated. Consequently, the problem is relinearized at x (1). The result is Minimize

11 8 ˜ ƒ(x; x (1)) ⫽ 2.284 ⫹ 2.444(x1 ⫺ –– 9 ) ⫹ 1.778(x2 ⫺ – 9)

11 8 Subject to g˜ 1(x; x (1)) ⫽ ⫹0.6049 ⫹ 2.444(x1 ⫺ –– 9 ) ⫺ (x2 ⫺ – 9) ⭓ 0 11 8 h˜ 1(x; x (1)) ⫽ ⫺0.0123 ⫺ (x1 ⫺ –– 9 ) ⫺ 1.778(x2 ⫺ – 9) ⫽ 0

–25 ⭓ x1 ⭓ –21

3 ⭓ x2 ⭓ 0

The new solution point lies at the intersection of g˜ 1(x; x (1)) ⫽ 0 and h˜ 1(x; x ) ⫽ 0 and is x (2) ⫽ (1.0187, 0.9965). The quality constraint is still violated [h1(x (2)) ⫽ ⫺0.0117], but to a lesser extent. If we continue in this fashion for two iterations, the solution x* ⫽ (1, 1) is reached with high accuracy. From the following tabulation of the successive function values (1)

Iteration

ƒ

g

h

0 1 3 Optimum

5 2.284 2.00015 2.00

3 0.605 3.44 ⫻ 10⫺4 0.0

⫺1.0 ⫺0.0123 ⫺1.18 ⫻ 10⫺5

0.0

it is evident that both the objective function and the equality constraint value improve monotonically. The inequality constraint g(x) remains feasible at all intermediate points. This type of monotonic behavior will occur if the problem functions are mildly nonlinear, as in the flatter portions of the parabolic constraints of Example 8.3. However, if more pronounced nonlinearities are encountered, then the monotonic improvement of the objective and constraint function values breaks down. Typically, the linearization of the objective function proves such a poor approximation that when the LP solution is used to evaluate the true objective function, a worse rather than an improved value is obtained. Alternatively, the linearization of the nonlinear constraints is such a poor approximation that when the LP solution is used to evaluate the actual con-

8.1


349

straints, the current LP solution is further outside the feasible region than its predecessor. If either or both of these situations occur during the iterations, the algorithm may fail to converge. This type of behavior is illustrated in the next example. Example 8.4

Suppose we begin the solution of the problem given in Example 8.3 at the point x 0 ⫽ (1.9999, 0.01). At this point, ƒ(x 0) ⫽ 3.9997 g1(x 0) ⫽ 3.9995 ⬎ 0 h1(x 0) ⫽ 0.0 That is, the point is feasible. The linearized subproblem in this instance becomes, after simplification, Minimize

⫺3.9997 ⫹ 3.9998x1 ⫹ 0.02x2

Subject to

3.9998x1 ⫺ x2 ⫺ 3.9996 ⭓ 0 ⫺x1 ⫺ 0.02x2 ⫹ 2.0001 ⫽ 0

–52 ⭓ x1 ⭓ –21

3 ⭓ x2 ⭓ 0

The linearized constraints are shown in Figure 8.5, and the solution to the linearized subproblem is x (1) ⫽ (1.9401, 3). In going from x 0 to x (1) the linearized objective function does actually improve from ƒ˜ (x 0; x 0) ⫽ 3.9997

Figure 8.5. Linearized problem feasible region, Example 8.4.

350


to ƒ˜ (x (1); x 0) ⫽ 3.8203. However, the actual objective function does not improve; that is, ƒ(x 0) ⫽ 3.9997

but ƒ(x (1)) ⫽ 12.764

Moreover, the new point turns out to be substantially infeasible, g1(x (1)) ⫽ 0.764 ⬎ 0 h1(x (1)) ⫽ ⫺8.9401 ⫽ 0 Thus, in this case, the iteration actually shows divergence. Step-Size Adjustment. From Example 8.4 it is evident that the linearization cannot be relied upon to yield uniformly good results. Although the linearization is always a good approximation within some suitably small neighborhood of any given base point, it can be spectacularly bad farther away from the point, especially if the function undergoes significant changes of slope within the neighborhood of the base point. One way to ensure that the linearization is used only within the immediate vicinity of the base point is to impose limits on the allowable increments in the variables, that is, for each subproblem constructed around base point x (t), we impose the bounds ⫺␦i ⭐ xi ⫺ x (t) i ⭐ ␦i

i ⫽ 1, . . . , N

where ␦i is some suitably chosen positive step-size parameter. This parameter is also referred to as defining a trust region, that is, a domain within which the linearization approximations are sufficiently accurate to be trusted. From a practical point of view this means that the parameters ␦i must be chosen small enough so that convergence is forced, that is, so that the original objective function is improved and also that the constraint infeasbilities are reduced in proceeding from x (t) to the solution x (t⫹1) of the current subproblem. Unfortunately, such a choice frequently leads to very slow overall convergence of the iterations. Example 8.5

Suppose the bounds ␦1 ⫽ ␦2 ⫽ 0.5 are imposed upon the subproblem generated in Example 8.4. This in effect adds the constraints 1.9999 ⫺ 0.5 ⭐ x1 ⭐ 0.5 ⫹ 1.9999 0.01 ⫺ 0.5 ⭐ x2 ⭐ 0.5 ⫹ 0.01 to the problem, which, as shown in Figure 8.6, amounts to constructing a hypercube around the base point x 0. The solution will now lie at the point x

8.1


351


at which the line h˜ (x; x 0) ⫽ 0 intersects the upper bound x2 ⫽ 0.51. The solution to the restricted subproblem is thus x ⫽ (1.9899, 0.51) with objective function value ƒ(x) ⫽ 4.2198 and constraint values g1(x) ⫽ 3.450 ⬎ 0

h1(x) ⫽ ⫺0.25 ⫽ 0

Note that although the constraint infeasibility is improved, the objective function value remains worse than at the starting point. This indicates that the ␦’s chosen are still too large and should be reduced further. It turns out, as can be verified by trial-and-error evaluations, that the ␦’s will have to be reduced to 0.05 before improvement is observed. With ␦1 ⫽ ␦2 ⫽ 0.05, the solution will be x ⫽ (1.9989, 0.05) g1(x) ⫽ 3.936 ⬎ 0

ƒ(x) ⫽ 3.9992 h1(x) ⫽ ⫺0.0025

But that in turn represents only a very small improvement over the base point x 0. The difficulties with this modification of the direct approach become rather obvious from this example. To achieve solutions that are feasible or near

352


feasible and at the same time yield improvement in the objective function value, the allowable increment may, in the presence of strong nonlinearities, have to be decreased to such an extent that only very slow progress in the iterations is achieved. Since the effort required to construct the linearized subproblems and to solve the resulting linear program is appreciable, the overall computational efficiency diminishes dramatically. An algorithm employing basically the solution strategy of Example 8.5 was first proposed by R. E. Griffith and R. A. Stewart of the Shell Oil Company [4]. Subsequent implementations were reported by H. V. Smith of IBM [5], by G. W. Graves of Aerospace Corporation ([6], cited in ref. 7), and others. The program developed by Griffith and Stewart, called MAP, includes special heuristics for adjusting the magnitudes of the ␦’s as the iterations progress. The IBM code, called P␾P II, adjusts the ␦’s by using estimates of the local linearization error, but this restricts its applicability to inequalityconstrained problems. Graves’s version, also applicable only in inequalityconstrained problems, differs from the other two primarily in incorporating estimates of the remainder term of the first-order Taylor expansion of the constraints into the linearization formulas. Moreover, a scaling parameter is introduced to regulate the size of the correction to the successive linear programming (SLP) solutions. Convergence of Graves’s algorithm can be proved, provided the constraints gj (x) ⭓ 0 are concave and the objective function is convex. Use of Penalty Functions. Observe that the step-bounding logic of these strategies can be viewed as a one-parameter search of some penalty function

P(x, p) ⫽ ƒ(x) ⫹ p⍀(g, h, x) over the direction defined by the vector x (t⫹1) ⫺ x (t). Thus, a conceptual twostep algorithm analogous to the Frank–Wolfe algorithm could be formulated. In the first step the LP is constructed and solved to yield a new point x. In the second step, for a suitable choice of parameter p, the line search problem Minimize P(x(␣), p) where x(␣) ⫽ x (t) ⫹ ␣(x ⫺ x (t))

␣⭓0

would be solved to yield the new point x (t⫹1) ⫽ x (t) ⫹ ␣ (t⫹1)(x ⫺ x (t)). Presumably some adjustment of the penalty parameter ought to be incorporated, and possibly the rigorous minimization of P could be relaxed to seeking ‘‘sufficient’’ improvement. Strategies of this general type were investigated by Palacios-Gomez et al. [8]. A representative version of such a successive approach would consist of

8.1


353

two primary phases as outlined below for a given specified tolerance ⑀ ⬎ 0 and step-size parameter ␦: Phase Check If at x (t) the exterior penalty term ⍀ satisfies ⍀(g, h, x) ⬎ ␧, then go to phase I. Otherwise go to phase II. Phase I Step 1. Linearize at x (t) and solve the LP problem to obtain x. Step 2. If ⍀(g, h, x) ⬍ ⍀(g, h, x (t)), then x (t⫹1) ⫽ x and proceed to the phase check. Otherwise reduce ␦ (e.g., ␦ ⫽ ␦ /2) and go to step 1. Phase II Step 1. Linearize and solve the LP problem to obtain x. Step 2. If P(p, x) ⬍ P(p, x (t)), set x (t⫹1) ⫽ x, increase the penalty parameter p, and go to the phase check. Otherwise reduce ␦ and go to step 1. Termination occurs when both 㛳x ⫺ x (t)㛳 and 兩P(p, x) ⫺ P(p, x (t))兩 are sufficiently small. Example 8.6

Suppose a Palacios-type strategy is applied to the problem of Example 8.3 at the point x (1) ⫽ (1.9899, 0.51) obtained in Example 8.5 with ␦ ⫽ –12 . For illustrative purposes we choose ␧ ⫽ 10⫺2, and ⍀(g, h, x) ⫽ h21 ⫹ [min(g1(x), 0)]2 1 ⫺2 At x (1), we have ⍀ ⫽ h21 ⫽ –– 16 ⬎ 10 . Therefore, the phase check directs the calculations to phase I.

Step 1. The LP problem constructed at x (1) is Minimize

3.9798⌬ x1 ⫹ 1.02⌬ x2

Subject to

3.45 ⫹ 3.9798⌬ x1 ⫺ ⌬ x2 ⭓ 0 ⫺0.25 ⫺ ⌬ x1 ⫺ 1.02⌬ x2 ⫽ 0 ⫺–12 ⭐ (⌬ x1, ⌬ x2) ⭐ –12

The solution is x ⫽ (1.4899, 0.7551).

354


Step 2. At x, g(x) ⫽ 1.4647 and h1(x) ⫽ ⫺6.007 ⫻ 10⫺2. Thus 1 ⍀(g, h, x) ⫽ 3.6 ⫻ 10⫺3 ⬍ ⍀(g, h, x (1)) ⫽ –– 16

Therefore, x (2) ⫽ x The phase check indicates that ⍀(g, h, x) ⬍ 10⫺2 and thus directs the calculations to phase II. Step 1. The LP problem at x (2) is Minimize Subject to

2.9798⌬ x1 ⫹ 1.5102⌬ x2 1.4647 ⫹ 2.9798⌬ x1 ⫺ ⌬ x2 ⭓ 0 ⫺6.007 ⫻ 10⫺2 ⫺ ⌬ x1 ⫺ 1.5102⌬ x2 ⫽ 0 ⫺–21 ⭐ (⌬ x1, ⌬ x2) ⭐ –12

The solution is x ⫽ (1.1369, 0.9889). Step 2. At this solution, ƒ(x) ⫽ 2.2703

g1(x) ⫽ 0.3036

h1(x) ⫽ ⫺0.1147

Hence, P(102, x) ⫽ 3.586 ⬎ P(102, x (2)) ⫽ 3.150. The step size ␦ must be reduced, say by –12 , and step 1 is repeated with ⫺–14 ⭐ (⌬ x1, ⌬ x2) ⭐ –14 . At the solution, x ⫽ (1.2399, 0.8809), the function values are ƒ(x) ⫽ 2.31327, g1(x) ⫽ 0.6565, and h1(x) ⫽ ⫺0.01582. Thus P(102, x) ⫽ 2.338 ⬍ P(102, x (2)) and we set x (3) ⫽ x, increase p to, say, 1000, and go to the phase check. At the cost of one more LP solution, the point x ⫽ (1.0215, 0.9958) is obtained with P(103, x) ⫽ 2.210 ⬍ P(102, x (3)) ⬍ P(103, x (3)) Thus, x (4) ⫽ x. With one further iteration, the very close optimal solution estimate x (5) ⫽ (1.00019, 0.99991) is calculated. Note that by making use of the penalty function and penalty term as measures of progress in the solution process, drastic reductions in the step-size bound ␦ were avoided and thus good progress could be made. Unfortunately,

8.1


355

although the use of the penalty function enhances the performance of the SLP strategy, it does not guarantee convergence in the general case. In fact, convergence results are available only in the linearly constrained case [8]. Of course, in all successive LP variants, if the solution of the LP approximation is such that x (t⫹1) ⫺ x (t) ⫽ 0, then it follows that there is no ⌬ x ⫽ 0 such that ⵜƒ 䡠 ⌬ x ⬍ 0 ⵜgj 䡠 ⌬ x ⬎ 0

j ⫽ 1, . . . , J

ⵜhk 䡠 ⌬ x ⫽ 0

k ⫽ 1, . . . , K

and consequently the Kuhn–Tucker conditions must hold [8]. A further refinement of the use of a penalty function together with a trust region construction was developed by Zhang et al. [9]. This version, called the penalty successive linear programming method, was shown to be convergent under mild conditions. However, the convergence rate is linear—analogous to the gradient method. 8.1.3

Discussion and Applications

In spite of the computational enhancements of the basic SLP strategy that are obtained by Palacios-Gomez et al. [8] and others [9, 10], the key limitation is the need to resort to the solution of a sequence of linear programs. If the problem is largely linear with only modest nonlinear contributions, then this overhead is acceptable. In this context, a good measure of nonlinearity is given by the number of nonlinear variables, that is, variables that appear in a nonlinear form in at least one problem function. For problems with a low relative number of nonlinear variables, the use of the SLP strategy has found wide acceptance in various industries. For instance, Lasdon and Warren [11] report on the use of this approach to solve large oil field development problems by both Gulf Oil and the Kuwait Oil Company. Problems involving up to 4392 linear and 400 nonlinear variables are reported to have been solved. Marathon Oil is reported to be using P␾P II as part of a hierarchical model for optimizing a complete refinery. Coefficients of the P␾P II model are obtained from up to 20 different process simulation models [11]. Similar specialized applications are reported by Chevron for solving refinery blending problems and by Union Carbide to solve multi-time-period process optimization problems. In the latter case, the intermediate linearized subproblems are reported to involve up to 10,000 variables. However, the number of nonlinear variables is small, so that the SLP subproblems are each solved in an average of only 23 simplex iterations [11]. Similar results are reported for Exxon applications by Baker and Lasdon [12]. Implementations of the SLP strategy are available (see, e.g., Xpress-SLP [13]). Additionally, most commercial LP packages do allow the LP solver to be callable from a C or

356


FORTRAN routine and provide utilities for updating the constraint set and/ or constraint coefficients. Thus, if an SLP implementation is not accessible, then penalty SLP strategies can readily be developed. In the next example, we present a representative but reduced-dimension refinery application involving a small number of nonlinear elements. Example 8.7

Pooling Problem

In refinery and other processing applications such as coal-blending plants, the need frequently arises to mix (or pool) several raw products to form an intermediate mixture, which is subsequently used as one of the components in a final blended product. The need for pooling usually occurs because of limited storage tank availability or because the raw products must be shipped in a single tank car, tanker, or pipeline. A simplified system requiring pooling is shown in Figure 8.7. In this system [14], two raw product streams with flows V1 and V2 and sulfur compositions 3 and 1 percent, respectively, must be pooled and then blended with a third raw product stream with flow F and 2 percent sulfur composition to produce two final blends. The first blend must have maximum sulfur composition of 2.5 percent and maximum demand of 100 bbl/hr; the second blend has maximum sulfur composition of 1.5 percent and maximum demand of 200 bbl/hr. The costs of the three raw streams are 6, 16, and 10 $/bbl, and the selling prices of the product blends are 9 and 10 $/bbl. Determine the flows and pool composition so as to maximize profit. Let Pi Fi Vi x

⫽ ⫽ ⫽ ⫽

flow flow flow pool

of blend i, bbl/hr of third raw product used in blend i of pooled intermediate used in blend i sulfur composition

From total flow balances,

Figure 8.7. Schematic of pooling problem, Example 8.7.

8.1


357

V1 ⫹ V2 ⫽ V1 ⫹ V2 F1 ⫹ V1 ⫽ P1 F2 ⫹ V2 ⫽ P2 From sulfur balances, 0.03V1 ⫹ 0.01V2 ⫽ x(V1 ⫹ V2) 0.02F1 ⫹ xV1 ⭐ 0.025P1 0.02F2 ⫹ xV2 ⭐ 0.015P2 In addition, the maximum demands reduce to the bounds P1 ⭐ 100

P2 ⭐ 200

and the pool composition must be bounded by the raw stream compositions, 0.01 ⭐ x ⭐ 0.03 The objective function in this case is simply the linear function 9P1 ⫹ 10P2 ⫺ 6V1 ⫺ 16V2 ⫺ 10F The optimization problem thus reduces to Maximize

ƒ ⫽ 9P1 ⫹ 10P2 ⫺ 6V1 ⫺ 16V2 ⫺ 10F

Subject to h1 ⫽ V1 ⫹ V2 ⫺ V1 ⫺ V2 ⫽ 0 h2 ⫽ 0.03V1 ⫹ 0.01V2 ⫺ x(V1 ⫹ V2) ⫽ 0 h3 ⫽ F1 ⫹ V1 ⫺ P1 ⫽ 0 h4 ⫽ F2 ⫹ V2 ⫺ P2 ⫽ 0 g1 ⫽ 0.025P1 ⫺ 0.02F1 ⫺ xV1 ⭓ 0 g2 ⫽ 0.015P2 ⫺ 0.02F2 ⫺ xV2 ⭓ 0 and P1 ⭐ 100

P2 ⭐ 200

0.01 ⭐ x ⭐ 0.03

358


Note that this problem would be an LP problem except for the product terms xV1 and xV2 that arise in h2, g1, and g2. If an initial guess were made of x, then the problem could be solved as a linear program. The solution could be used to update x, and the process would be continued until convergence was obtained. However, as observed by Haverly [14], such a recursive solution procedure is not reliable. Since equality constraint h2 is nonlinear, the feasible region for this problem is nonconvex; hence, the problem admits multiple local solutions. As pointed out by Lasdon et al. [15], this problem has three Kuhn–Tucker points: 1. A saddlepoint with x ⫽ 0.02 and all flows zero 2. A local maximum with x ⫽ 0.03, P1 ⫽ 100, F ⫽ V1 ⫽ 50, all other flows zero, and a profit of $100/hr 3. A global maximum with x ⫽ 0.01, V2 ⫽ F2 ⫽ 100, P2 ⫽ 200, all other flows zero, and a profit of $400/hr The recursive solution strategy leads to one of these three points depending upon the initial value of x selected. However, as reported by Lasdon et al. [15], solution using the SLP strategy devised by Palacios-Gomez et al. [8] leads to the global maximum regardless of the initial value of x selected, requiring between two and three LP subproblem solutions. It is not clear why the SLP strategy avoids the local solutions; however, it certainly is a welcome result. Since in pooling applications of practical scope the number of pools, raw streams, product blends, and blend characteristics or compositions would be much larger, the possibility of local maxima increases, and hence the use of SLP is clearly to be preferred. In summary, the SLP strategy, particularly with the penalty function modifications is a useful and practical approach for solving large NLP problems with a moderate degree of nonlinearity. For highly nonlinear problems, however, this solution strategy will require substantial reductions in the step size in order to ensure progress in the search, thus resulting in slow convergence. In fact, the typically slow convergence with highly nonlinear problems has in the past led to the use of the name small step gradient methods for this family of NLP techniques. Improvements in the direct linear approximation methods have taken three alternative forms: 1. Global approximation of the problem nonlinearities by means of piecewise linear functions 2. Accumulation of successive linearizations from iteration to iteration so as to better define the feasible region 3. Use of the local linear approximation to generate only a direction for movement rather than a complete step

8.2

SEPARABLE PROGRAMMING

359

In the following sections we discuss the ideas involved in each of these constructions. As will become apparent, the first two approaches are applicable to only a limited class of problems, while the third has led to very efficient and powerful general-purpose algorithms.

8.2


Separable programming is a technique first developed by C. E. Miller [16] in 1963 by means of which certain types of nonlinear constrained optimization problems can be reformulated into equivalent problems involving only linear functions. The resulting approximating problems are then solved using a specially modified simplex algorithm. The motivation for this technique stems from the observation that a good way of improving the linear approximation to a general nonlinear function over a large interval is to partition the interval into subintervals and construct individual linear approximations over each subinterval. This is known as constructing a piecewise linear approximation. In this section, we first consider the use of this construction for singlevariable functions. Then we extend the construction to multivariable functions. We also examine the modifications to standard LP methodology that must be made to effectively solve the resulting linear subproblems. Finally, we discuss some practical applications of the separable programming construction. 8.2.1

Single-Variable Functions

In the special case of single-variable functions, the development of a piecewise linear approximation is quite straightforward. Consider an arbitrary continuous function ƒ(x) of a single variable defined over an interval 0 ⭐ x ⭐ a. Suppose we arbitrarily choose a grid of K points spaced over the interval of interest and denote the points x (k), k ⫽ 1, . . . , K. For each x (k), we evaluate ƒk ⫽ ƒ(x (k)), and between every pair of adjacent grid points x (k), x (k⫹1) we draw a straight line connecting the points (x k, ƒk) and (x k⫹1, ƒk⫹1). As shown in Figure 8.8, we will have formed an approximating function ƒ˜ (x) that is a piecewise linear function. The equation of the line connecting points (x (k), ƒ(k)) and (x (k⫹1), ƒ(k⫹1)) is given by ˜ ƒ(x) ⫽ ƒ(k) ⫹

ƒ(k⫹1) ⫺ ƒ(k) (x ⫺ x (k)) x (k⫹1) ⫺ x (k)

(8.1)

An equation of this form will be associated with each subinterval x (k) ⭐ x ⭐ x (k⫹1). Consequently, the entire approximation to ƒ(x) will consists of K ⫺ 1

360


Figure 8.8. Piecewise linear approximation.

such linear equations, with each one applicable over its restricted subinterval of validity. It is convenient to condense this collection of linear equations into a more general form. First of all, observe that any x in the range x (k) ⭐ x ⭐ x (k⫹1) can be written as x ⫽ ␭(k)x (k) ⫹ ␭(k⫹1)x (k⫹1)

(8.2)

where ␭(k) and ␭(k⫹1) are two nonnegative parameters normalized to one, that is, ␭(k) ⫹ ␭(k⫹1) ⫽ 1

Suppose we substitute this expression for x into Eq. (8.1): ˜ ƒ(x) ⫽ ƒ(x (k)) ⫹

ƒ(k⫹1) ⫺ ƒ(k) (k⫹1) (k⫹1) [␭ x ⫺ (1 ⫺ ␭(k))x (k)] x (k⫹1) ⫺ x (k)

Since 1 ⫺ ␭(k) ⫽ ␭(k⫹1), it follows that ˜ ƒ(x) ⫽ ƒ(x (k)) ⫹ ␭(k⫹1)(ƒ(k⫹1) ⫺ ƒ(k)) ⫽ ␭(k)ƒ(k) ⫹ ␭(k⫹1)ƒ(k⫹1)

(8.3)

Thus in the interval x (1) ⭐ x ⭐ x (K), each x and its approximate function ˜ value ƒ(x) can be generated by assigning appropriate values to the two ␭ variables that correspond to the subinterval within which x lies. Since any given value of x can lie in only one subinterval, it further follows that only

8.2


361

the two ␭’s associated with that subinterval will have nonzero values. The other ␭’s must all be equal to zero. Consequently, the K ⫺ 1 sets of Eqs. (8.2) and (8.3) can be combined conveniently into the two equations

冘␭ K

x⫽

冘␭ K

˜ ƒ(x) ⫽

(k) (k)

x

k⫽1

(k) (k)

ƒ

(8.4)

k⫽1

where

冘␭ K

(k)

(i)

⫽1

k⫽1

(ii)

␭(k) ⭓ 0

k ⫽ 1, . . . , K

and (iii)

␭(i)␭(j) ⫽ 0

if j ⬎ i ⫹ 1; i ⫽ 1, K ⫺ 1

Condition (iii) merely requires that no more than two of the ␭(i)’s be positive and if two are positive, say ␭(j) and ␭(i) with j ⬎ i, then it must be true that j ⫽ i ⫹ 1, that is, the ␭’s must be adjacent. This restriction ensures that only points lying on the piecewise linear segments are considered part of the approximating function. To see this, note from Figure 8.8 that if ␭(1) and ␭(2) were positive and ␭(3) to ␭(6) were zero, then the resulting point would lie on the line segment joining points A and B, a portion of the approximating function. However, if, say, ␭(1) and ␭(3) were allowed to be positive and the other ␭’s were all equal to zero, then a point on the line connecting A and C, which is not part of the approximating function, would be generated. Finally if, say, ␭(3) alone were positive, then from condition (i) it would have to be equal to 1, and the point C lying on both ƒ(x) and ƒ˜ (x) would be generated. The piecewise linear construction does require an increase in problem size. In Figure 8.8 the nonlinear function ƒ(x) is converted into a linear function ƒ˜ (x) at the cost of introducing six new variables (the ␭’s) and two linear side conditions (8.4) and (i). Moreover, if we wanted a closer approximation of ƒ(x), we could introduce additional grid points; however, this would introduce additional variables ␭(k). The accuracy of the approximation is also strongly dependent on a judicious selection of the grid-point locations. The best way to select grid-point locations is to work from a plot of the function over the variable range of interest. Grid points should normally be located in the vicinity of inflection points (such as C in Figure 8.8) and should be concentrated in the regions in which the function changes values rapidly, that is, regions with large function derivative values. This means that in general the grid-point locations will be spaced nonuniformly.

362


8.2.2

Multivariable Separable Functions

The foregoing construction can be generalized to multivariable functions only if the multivariable functions are separable. Definition A function ƒ(x) of the N-component vector variable x is said to be separable if it can be expressed as the sum of single-variable functions that each involve only one of the N variables. Thus,

冘 ƒ (x ) N

ƒ(x) ⫽

i

i

i⫽1

For example, the function ƒ(x) ⫽ x 21 ⫹ ex2 ⫹ x 22 / 3 is separable, but ƒ(x) ⫽ x1sin(x2 ⫹ x3) ⫹ x2ex3 is not. The restriction of the approximation technique to separable functions may seem to be quite restrictive and it is. However, the class of separable nonlinear functions is nonetheless quite large. Moreover, there are various devices by means of which some functions that are not directly separable can be rewritten in separable form. For instance, the product x1x2 can be rewritten as x 23 ⫺ x 24 with the side conditions x3 ⫽ –21 (x1 ⫹ x2)

x4 ⫽ –21 (x1 ⫺ x2)

and

(8.5)

The generalized product

写x N

c

ai i

i⫽1

can be replaced by y with the side conditions

冘 a ln(x ) N

ln(y) ⫽ ln(c) ⫹

i

i⫽1

provided the xi are restricted to be positive.

i

(8.6)

8.2


363

Under the restriction that the multivariable function ƒ(x) is separable and continuous, the construction of a piecewise linear approximation to ƒ(x) proceeds as follows: Subdivide the interval of values on each variable xi with Ki grid points. Thus, each variable will be associated with a sequence of points (Ki) x (L) ⫽ x (1) ⬍ x (2) ⬍ 䡠䡠䡠 ⬍ x (j) ⫽ x (U) i i i i ⬍ 䡠䡠䡠 ⬍ xi i

where x (L) and x i(U) are the given or assumed values of the lower and upper i bounds on variable xi. ⫽ ƒi(x (k) If we adopt the shorthand notation ƒ(k) i i ), then the approximating function ƒ˜ (x) will be given by

冘 K1

˜ ƒ(x) ⫽

冘 K2

(k) ␭(k) 1 ƒ1 ⫹

k⫽1

冘␭ KN

␭2(k)ƒ2(k) ⫹ 䡠䡠䡠 ⫹

k⫽1

(k) (k) N N

ƒ

k⫽1

where

冘␭ Ki

xi ⫽

i ⫽ 1, . . . , N

(k) (k) i i

x

k⫽1

given that for i ⫽ 1, . . . , N

冘␭ Ki

(k) i

(i)

⫽1

k⫽1

␭(k) i ⭓ 0

(ii) (iii)

(j) ␭(i) i ␭i ⫽ 0

k ⫽ 1, . . . , Ki

if j ⬎ i ⫹ 1; i ⫽ 1, 2, . . . , Ki ⫺ 1

In this manner each nonlinear function ƒ(x) is replaced by a function ƒ˜ (x) that is linear in the variables. Of course, if a very accurate approximation to ƒ(x) is required, then the number of grid points and hence variables ␭ can become quite large. Again, linearity is attained at the price of substantially increased dimensionality. Example 8.8

Consider the function ƒ(x) ⫽ ƒ1(x1) ⫹ ƒ2(x2) ⫽ (1 ⫺ (1 ⫺ x1)3) ⫹ (0.1x 24) over the hypercube 0 ⭐ x1 ⭐ 2 and 0 ⭐ x2 ⭐ 3. Let the approximating function to ƒ1(x1) be constructed by subdividing the interval 0 ⭐ x1 ⭐ 2 into three sections by using the four grid points 0, 0.5, 1.5, and 2.0. Thus,

364


(2) (3) (4) x1 ⫽ 0␭(1) 1 ⫹ 0.5␭1 ⫹ 1.5␭1 ⫹ 2.0␭1

and ƒ˜ 1(x1) ⫽ ƒ1(0)␭1(1) ⫹ ƒ1(0.5)␭1(2) ⫹ ƒ1(1.5)␭1(3) ⫹ ƒ1(2.0)␭1(4) (2) (3) (4) ⫽ 0␭(1) 1 ⫹ 0.875␭1 ⫹ 1.125␭1 ⫹ 2.0␭1

Similarly, to construct the approximation to ƒ2(x2), suppose we choose five grid points 0, 1, 2, 2.6, and 3 with corresponding ƒ2 values of 0, 0.1, 1.6, 4.57, and 8.1. Then, ˜ 1, x2) ⫽ 0␭1(1) ⫹ 0.875␭1(2) ⫹ 1.125␭1(3) ⫹ 2.0␭1(4) ⫹ 0␭2(1) ƒ(x (3) (4) (5) ⫹ 0.1␭(2) 2 ⫹ 1.6␭2 ⫹ 4.57␭2 ⫹ 8.1␭2

with (2) (3) (4) x1 ⫽ 0␭(1) 1 ⫹ 0.5␭1 ⫹ 1.5␭1 ⫹ 2.0␭1 (2) (3) (4) (5) x2 ⫽ 0␭(1) 2 ⫹ 1␭2 ⫹ 2␭2 ⫹ 2.6␭2 ⫹ 3␭2

The ␭(k) are restricted to satisfy i

冘␭ 4

(k) 1

k⫽1

冘␭ 5

⫽1

(k) 2

⫽1

␭i(k) ⭓ 0

k⫽1

(k) and only two adjacent ␭(k) 1 ’s and two adjacent ␭2 ’s can be positive. For example,

␭(2) 1 ⫽ 0.4

␭1(3) ⫽ 0.6

␭1(1) ⫽ ␭1(4) ⫽ 0

␭(3) 2 ⫽ 0.5

␭2(4) ⫽ 0.5

␭2(1) ⫽ ␭2(2) ⫽ ␭2(5) ⫽ 0

will generate the point x ⫽ (1.1, 2.3) and the approximate function value ƒ˜ (1.1, 2.3) ⫽ 4.11. The exact function value is ƒ(1.1, 2.3) ⫽ 3.8. 8.2.3

Linear Programming Solutions of Separable Problems

By using piecewise linear approximations of the form shown in Example 8.8, any NLP problem in which both the objective function and the constraints are separable can be reformulated as a large linear program in the variables ␭ that can be solved by conventional LP codes. The only feature requiring special attention is condition (iii). this restriction can, however, be readily accommodated, since in the ordinary simplex method the basic variables are the only ones that can be nonzero. Thus, prior to entering one of the ␭’s into

8.2


365

the basis (which will make it nonzero), a check is made to ensure that no more than one other ␭ associated with the corresponding variable xi is in the basis (is nonzero) and, if it is, that the ␭’s are adjacent. If these checks are not satisfied, then the ␭ to be entered into the basis is rejected and another is selected. This modification to the normal simplex rules required to ensure that condition (iii) is always met is known as restricted basis entry. Recall that the reason for imposing condition (iii) and for restricted basis entry is to ensure that the points generated lie on the piecewise linear segments rather than between them. In the following example, we illustrate the use of separable programming constructions to solve a nonlinear program via an approximating linear program. Example 8.9

Maximize

ƒ(x) ⫽ x 41 ⫹ x2

Subject to g1(x) ⫽ 9 ⫺ 2x 21 ⫺ 3x2 ⭓ 0 x1 ⭓ 0

x2 ⭓ 0

Since both problem functions are separable, we can immediately identify ƒ(x) ⫽ ƒ1(x1) ⫹ ƒ2(x2) g1(x) ⫽ g11(x1) ⫹ g12(x2) where ƒ1(x1) ⫽ x 14 g11(x1) ⫽ ⫺2x 21

ƒ2(x2) ⫽ x2 g12(x2) ⫽ ⫺3x2

Since ƒ2 and g12 are linear functions, they need not be replaced by piecewise linear functions. However, the other two, ƒ1 and g11, will need to be approximated. Suppose we choose to construct our approximation over the interval 0 ⭐ x1 ⭐ 3, and suppose we use four equidistant grid points for this purpose. The data required for the piecewise approximation are tabulated below: k

x (k) 1

ƒ1(x (k) 1 )

g11(x 1(k))

1 2 3 4

0 1 2 3

0 1 16 81

⫺2 ⫺8 ⫺18

0

366


Thus the approximation to ƒ1(x1) becomes ƒ˜ 1(x1) ⫽ (0)␭1(1) ⫹ (1)␭1(2) ⫹ (16)␭1(3) ⫹ (81)␭1(4) and the approximation to g11(x1) becomes (2) (3) (4) g˜ 11(x1) ⫽ (0)␭(1) 1 ⫺ 2␭1 ⫺ 8␭1 ⫺ 18␭1

with the restriction (2) (3) (4) ␭(1) 1 ⫹ ␭1 ⫹ ␭1 ⫹ ␭1 ⫽ 1

The resulting linear approximating problem thus becomes Maximize

(3) (4) ƒ(x) ⫽ ␭(2) 1 ⫹ 16␭1 ⫹ 81␭1 ⫹ x2

Subject to

2␭1(2) ⫹ 8␭1(3) ⫹ 18␭1(4) ⫹ 3x2 ⫹ x3 ⫽ 9 ␭1(1) ⫹ ␭1(2) ⫹ ␭1(3) ⫹ ␭1(4) ⫽ 1

all ␭1(k), x2, x3 ⭐ 0 Variable x3 is a slack variable inserted to convert the constraint to an equality. Suppose we convert the LP problem to tableau form and execute the usual simplex calculations. CJ 1 ␭

(2) 1

16

81

1

0

0

(3) 1

␭

(4) 1

␭

x2

x3

␭(1) 1

8

CB

Basis

0 0

x3 ␭(1) 1

2 1

①

18 1

3 0

1 0

0 1

9 1

CJ

1

16

81

1

0

0

0

↑

From the CJ row, we see that ␭1(4) is the entering variable, and from the minimum-ratio rule it follows that the outgoing variable is x3. But ␭(4) 1 cannot (1) (4) be brought into the basis, since ␭(4) 1 and ␭1 are not adjacent. Therefore, ␭1 is rejected and the next best value in the CJ row is sought. From the CJ row ␭(3) 1 is the next candidate to enter the basis, and from the minimum-ratio rule, ␭(1) 1 will leave the basis. Continuing in this fashion, we eventually achieve the optimal solution with (4) basic variables ␭(3) 1 ⫽ 0.9 and ␭1 ⫽ 0.1. The corresponding optimal value of x1 is

8.2

367


x1 ⫽ 0(0) ⫹ 1(0) ⫹ 2(0.9) ⫹ 3(0.1) ⫽ 2.1 Thus the optimal solution to the separable program is x˜ * ⫽ (2.1, 0) with ƒ˜ (x*) ⫽ 22.5. As can be readily verified, the exact optimum solution is x* ⫽ (2.12, 0)

ƒ(x*) ⫽ 20.25

Note that if, in the tableau, the restricted basis entry rule had been neglected, then ␭(4) 1 would have entered the basis and x3 would have left. The resulting next tableau would then have been CJ 1 ␭

(2) 1

16

81

1

0

0

␭

␭

x2

x3

x (1) 1

(3) 1

(4) 1

CB

Basis

81 0

␭(4) 1 ␭(1) 1

–19 –89

–49 –59

1 0

–16 ⫺–16

1 –– 18 1 ⫺–– 18

0 1

–12 –12

CJ

⫺8

⫺20

0

⫺12.5

⫺4.5

0

40.5

81 Apparently the optimal solution is ␭(1) ⫽ ␭1(4) ⫽ –21 , with ƒ˜ ⫽ –– 1 2 ⫽ 40.5. However, this is clearly not correct, since these values of the ␭’s give

x1 ⫽ 0(–21 ) ⫹ 1(0) ⫹ 2(0) ⫹ 3(–21 ) ⫽ 1.5 and at the point (1.5, 0) the actual objective function value is ƒ(1.5, 0) ⫽ (1.5)4 ⫹ 0 ⫽ 5.0625 Neglect of the basis entry rule can therefore lead to points that do not lie on the approximating segments and hence may yield meaningless LP solutions. For a special class of problems the restricted basis rule can, however, be relaxed. As proven in reference 17 (Theorem 11.3.1), given a separable nonlinear program of the form

冘 ƒ (x ) 冘 g (x ) ⭐ b N

Minimize

i

i

i⫽1 N

Subject to

ji

i

j

j ⫽ 1, . . . , J

i⫽1

xi ⭓ 0

i ⫽ 1, . . . , N

the restricted basis rule will always be satisfied at the optimum of the separable programming LP provided:

368


1. For all i, ƒi(xi) is either strictly convex or it is linear. 2. For all i and j, gji is either concave or it is linear. Since the conditions involve the convexity of the single-variable functions, they are easily checked and quite useful in practice.

8.2.4

Discussion and Applications

The solution obtained via separable programming is only as accurate as the underlying approximating problem. In fact, only if the components of the separable functions gji are all either linear or concave can it be guaranteed that the solution will even be a feasible solution of the original problem. Solutions of improved feasibility and accuracy can be generated by following either of two strategies: 1. Increase the overall accuracy of each approximating function to the desired level by adding more grid points. However, this increases the number of variables and thus, because of the restricted basis requirement, considerably increases the number of iterations required to solve the approximating problem. 2. Increase the accuracy of each approximating function locally in an iterative fashion. The problem is solved initially via a rather coarse grid. The grid is then refined only in the neighborhood of the resulting solution. The problem is solved again, and the refinement process continues until a sufficiently accurate solution is attained. In this manner the number of grid points can be kept constant at the price of repeated solution of the intermediate LP problems. The latter strategy is usually preferable. However, it should be noted that local iterative refinement can, in some cases, lead to convergence to false optima and hence must be used with care. In general, separable programming solutions will be only local minima. Attainment of global minima is guaranteed only in the case of convex objective functions and convex feasible regions. Clearly, separable programming is not suitable for general NLP problems. It is appropriate for applications that are largely linear with some nonlinear elements that can be expressed as separable functions. If highly accurate solutions are not required, then, by use of the piecewise linear construction, the approximate optimum will be obtained in a single solution of the LP subproblem. This feature has led some practitioners to rank separable programming very highly as a useful and practical tool [18]. Separable programming has found numerous applications in refinery scheduling and planning models [19]. We will consider one such application in the next example.

8.2

Example 8.10


369

Computing Tetraethyllead Requirements [19, 20]

One very important nonlinearity encountered in refinery modeling arises in computing the amount of tetraethyllead (TEL) that must be added to a gasoline blend to increase the gasoline octane number to required levels. The octane number is not a linear function of TEL. Instead, the exact relationship is defined using TEL susceptibility charts, such as Figure 8.9. The assumption used in the industry is that the octane number of a blend can be estimated as the volumetric average of the octane numbers of its components all measured at the same TEL concentration as that of the blend. Suppose N components plus TEL are to be blended to yield a gasoline with specified minimum research and motor octane number. The TEL concentration of the gasoline is not to exceed 3 cm3 TEL per U.S. gallon. For each of the components, the relationship between the octane numbers and the TEL concentration will be given by a chart similar to Figure 8.9. ⫽ ⫽ ⫽ ⫽

flow of component n, gal/day flow of blended gasoline, gal/day concentration of TEL in blend, cm3 /gal research octane number of component n when it contains TEL at concentration z mn(z) ⫽ motor octane number of component n when it contains TEL at concentration z r, m ⫽ minimum acceptable research (motor) octane number of blend w ⫽ total usage of TEL, cm3 /day

Let xn y z rn(z)

Figure 8.9. TEL susceptibility chart, Example 8.10.

370


From a flow balance,

冘x ⫽y n

n

Using the octane number blending assumption,

冘 r (z)x ⭓ ry 冘 m (z)x ⭓ my n

n

n

n

n

(8.7)

n

Finally, w ⫽ zy, where w would presumably be used in the objective function. Note that since rn and mn are functions of z, the two blending constraints are actually nonlinear functions. However, by choosing, say, four grid points, the curves of Figure 8.9 can be approximated by piecewise linear functions, as shown in Figure 8.10 and given in equation form below:

冘␭ ⫽1 4

z ⫽ a1␭1 ⫹ a2␭2 ⫹ a3␭3 ⫹ a4␭4

i

(8.8)

i⫽1

rn(z) ⫽ rn1␭1 ⫹ rn2␭2 ⫹ rn2␭3 ⫹ rn4␭4 mn(z) ⫽ mn1␭1 ⫹ mn2␭2 ⫹ mn3␭3 ⫹ mn4␭4 The blending constraint thus takes the form

Figure 8.10. Piecewise linear approximation to susceptibility chart.

(8.9)

8.2


冘 x (r ␭ ⫹ r ␭ ⫹ r ␭ ⫹ r ␭ ) ⭓ ry

371

N

n

n1 1

n2 2

n3 3

n4 4

(8.10)

n⫽1

which, however, includes the nonlinear variable products xn␭i. Let r (u) ⫽ maxn,i(rni) and y (u) be some reasonable upper bound on y, the gasoline production. Then, using substitution (8.5) discussed in Section 8.2, we can define ui ⫽ vi ⫽

冋冋

1 1 (u) (u) 2 r y 1 1 2 r (u)y (u)

冘 r x ⫹ ␭册冘 r x ⫺ ␭册 N

ni n

i

ni n

i

i ⫽ 1, . . . , 4

n⫽1 N

(8.11)

n⫽1

Note that since 0 ⭐ ␭i ⭐ 1

0⭐

and

1 r (u)y (u)

冘r x ⭐1 ni n

n

it follows that 0 ⭐ ui ⭐ 1

⫺–21 ⭐ vi ⭐ –21

(8.12)

Thus, the blending constraint (8.10) takes the separable form

冘 u ⫺ 冘 v ⭓ r ryy 4

4

2 i

2 i

i⫽1

i⫽1

(u) (u)

(8.13)

Finally, each ui, vi and its square is approximated by piecewise linear functions over the range given by Eqs. (8.12). For instance, with five equidistant grid points, ui ⫽ 0␤i1 ⫹ –41 ␤i2 ⫹ –21 ␤i3 ⫹ –43 ␤i4 ⫹ ␤i5

(8.14)

1 1 9 u2i ⫽ 0␤i1 ⫹ –– 16 ␤i2 ⫹ – 4 ␤i3 ⫹ –– 16 ␤i4 ⫹ ␤i5

(8.15)

and

冘␤ 5

ij

⫽1

j⫽1

Similar variables ␥ij would be required for the vi’s.

(8.16)

372


Thus, by using four grid points for the TEL concentration and five grid points for each ui, vi, we have introduced the following variables: ui, vi, ␭i

i ⫽ 1, . . . , 4

␤ij, ␥ij

i ⫽ 1, . . . , 4; j ⫽ 1, . . . , 5

or a total of 52. We have introduced the following equations: Eq. (8.8) and the normalization condition on the ␭’s, N equations of type (8.9), 8 equations of type (8.11), and 24 equations of the type (8.14)–(8.16). The total is 34 ⫹ N equations plus bounds (8.12). This may well seem like an enormous penalty to pay for adapting the octane number constraints (8.7) for LP solution. However, since these constraints normally constitute only a very small portion of a large otherwise linear refinery model, the grid-point constructions prove quite practical and are widely used in the industry [19, 20].

8.3

SUMMARY

This chapter explored two basic strategies for employing linear approximations in solving nonlinearly constrained problems: the direct SLP approach and global approximation via piecewise linear functions. In the direct linearization approach we found that the LP solution can be used only to define a direction that must be searched using the methods of Chapter 2. The Frank– Wolfe algorithm, applicable to linearly constrained problems, can thus be viewed as an alternating series of LP and line search subproblems. The application of the SLP approach to nonlinear constraints introduced the need to adjust the LP solution to maintain constraint feasibility. The Griffith and Stewart method of bounding the step to maintain both feasibility and descent leads to very slow progress; the use of a penalty function is much preferred. Nonetheless, even with these enhancements, we conclude that the SLP approach is most appropriate for problems with few nonlinear variables. The separable programming technique involves the use of piecewise linear constructions to approximate separable nonlinear functions over the full range of the problem variables. This construction does substantially increase problem dimensionality and normally requires grid readjustment if more accurate solutions are required. For nonconvex feasible regions, the resulting linear program must be solved with restrictions in the choice of basic variables to ensure that only points on the approximating segments are generated. The separable programming construction is most appropriate for problems that would be LP problems except for the presence of a few separable nonlinear elements. Such situations occur widely in planning and scheduling applications.

REFERENCES

373

REFERENCES 1. Frank, N., and P. Wolfe, ‘‘An Algorithm for Quadratic Programming,’’ Naval Res. Log. Quart., 3, 95–110 (1956). 2. Zangwill, W. I., Nonlinear Programming, Prentice-Hall, Englewood Cliffs, NJ, 1969. 3. Wolfe, P., ‘‘Convergence Theory in Nonlinear Programming,’’ in Integer and Nonlinear Programming (J. Abadie, Ed.), North-Holland, Amsterdam, 1970. 4. Griffith, R. E., and R. A. Stewart, ‘‘A Nonlinear Programming Technique for Optimization of Continuous Processing Systems,’’ Manag. Sci., 7, 379–391 (1961). 5. Smith, H. V., ‘‘A Process Optimization Program for Nonlinear Systems, Pop II,’’ 7090 H9 IBM UD21, SHARE General Program Library, 1965. 6. Graves, G. W., ‘‘Development and Testing of a Nonlinear Programming Algorithm,’’ Rep. No. ATR-64 (7040)-2, Aerospace Corporation, June 1964. 7. Clasen, R. J., G. W. Graves, and J. Y. Lu, ‘‘Sortie Allocation by a Nonlinear Programming Model for Determining a Munitions Mix,’’ Rep R-1411-DDPAE, Rand Corporation, Mar. 1974. 8. Palacios-Gomez, F. E., L. Lasdon, and M. Engquist, ‘‘Nonlinear Optimization by Successive Linear Programming,’’ Manag. Sci., 28(10), 1106–1120 (1982). 9. Zhang, J., N. H. Kim, and L. Lasdon, ‘‘An Improved Successive Linear Programming Algorithm,’’ Manag. Sci., 31(10), 1312–1331 (1985). 10. Beale, E. M. L., ‘‘Nonlinear Programming Using a General Mathematical Programming System,’’ in Design and Implementation of Optimization Software (H. J. Greenberg, Ed.), Sijthoff and Noordhoff, Netherlands, 1978. 11. Lasdon, L. S., and A. D. Warren, ‘‘Survey of Nonlinear Programming Applications,’’ Oper. Res., 28, 1029–1073 (1980). 12. Baker, T. E., and L. Lasdon, ‘‘Successive Linear Programming at Exxon,’’ Manag. Sci., 31(3), 264–274 (1985). 13. Dash Optimization, Englewood Cliffs, NJ, www.dashoptimization.com. 14. Haverly, C. A., ‘‘Studies of the Behavior of Recursion for the Pooling Problem,’’ SIGMAP Bull., 25, 19–28 (Dec. 1978). 15. Lasdon, L. S., A. D. Warren, S. Sarkar, and F. E. Palacios-Gomez, ‘‘Solving the Pooling Problem Using Generalized Reduced Gradient and Successive Linear Programming Algorithms,’’ SIGMAP Bull., 77, 9–15 (July 1979). 16. Miller, C. E., ‘‘The Simplex Method for Local Separable Programming,’’ in Recent Advances in Mathematical Programming (R. Graves and P. Wolfe, Eds.), McGrawHill, New York, 1963. 17. Bazaraa, M. S., H. D. Sherali, and C. M. Shetty, Nonlinear Programming: Theory and Algorithms, 2nd ed., Wiley, New York, 1993. 18. Beale, E. M. L., ‘‘Advanced Algorithmic Features for General Mathematical Programming Systems,’’ in Integer and Nonlinear Programming (J. Abadie, Ed.), North-Holland, Amsterdam, 1970. 19. El Agizy, ‘‘Applications of Separable Programming to Refinery Models,’’ ORSA 36th National Meeting, Nov. 1969. 20. Kawaratoni, T. K., R. J. Ullman, and G. B. Dantzig, ‘‘Computing Tetraethyl-lead Requirements in a Linear Programming Format,’’ Oper. Res., 8, 24–29 (1960).

374


PROBLEMS

8.1. Why is a line search necessary in the Frank–Wolfe algorithm? Why must the algorithm be initiated with a feasible point? 8.2. What is the fundamental reason for placing bounds on the changes in the variables in the Griffith/Stewart direct linearization method? 8.3. Suggest a method by means of which a feasible starting point can be computed for the Frank–Wolfe algorithm. 8.4. Since the LP subproblem of the Frank–Wolfe algorithm always yields corner points as a solution, explain how the algorithm would approach a point that is strictly in the interior of the feasible region. 8.5. Recall that in the SLP strategy using the penalty function, only a rough minimization of the penalty function is carried out. Discuss the merits of also solving the LP subproblem only approximately. Do you think just one LP iteration per subproblem might suffice? Explain. 8.6. Outline a penalty-type SLP strategy that would exclude linear constraints and variable bounds from the penalty function construction. Discuss the merits of such a strategy. 8.7. Under what problem conditions is separable programming likely to be preferable to SLP methods? 8.8. What is restricted basis entry, why is it necessary, and under what conditions can it be disregarded? 8.9. If a grid is selected for a given variable xn and used to approximate a function ƒ(xn), must the same grid be used for all other functions involving xn? Explain. 8.10. In the SLP strategy, is the Griffith–Stewart step-size reduction strategy more conservative than the use of a penalty function to select the step size? Why? Why not? 8.11. Consider the problem Minimize

ƒ(x) ⫽ –13 (x1 ⫹ 3)3 ⫹ x2

Subject to

3 ⭓ x1 ⭓ 1 2 ⭓ x2 ⭓ 0

(a) Construct a piecewise linear approximation to the cubic term over the range of x1 using four grid points. (b) Will restricted basis entry be necessary in solving the resulting LP problem? Explain.

PROBLEMS

375

(c) Solve the approximating linear program. Compare the solution to the actual minimum. 8.12. Consider the problem Minimize

ƒ(x) ⫽ (x1 ⫺ 2)2 ⫹ (x2 ⫺ 2)2

Subject to g1(x) ⫽ 3 ⫺ x1 ⫺ x2 ⭓ 0 g2(x) ⫽ 10x1 ⫺ x2 ⫺ 2 ⭓ 0 x2 ⭓ 0 (a) Can x 0 ⫽ (0, 0) be used as starting point for the Frank–Wolfe algorithm? Why? (b) Suppose x0 ⫽ (0.2, 0) is used. Solve the problem using the Frank– Wolfe algorithm. 8.13. Consider the problem Maximize

ƒ(x) ⫽ 2x1 ⫺ 3(x2 ⫺ 4)2

Subject to g1(x) ⫽ 10 ⫺ x 12 ⫺ x 22 ⭓ 0 g2(x) ⫽ 9 ⫺ x 12 ⫺ (x2 ⫺ 4)2 ⭓ 0 x1 ⭓ 0 (a) Given the starting point x 0 ⫽ (2, 2), calculate the next iteration point that would result if a small step gradient method were applied with ␦ ⫽ 1. (b) Repeat the calculation with the point x 0 ⫽ (0, 1) and with the point (1, 3). 8.14. Replace the objective function of Example 8.11 with the new objective function, max ƒ(x) ⫽ x 11 / 4 ⫹ x2 (a) Show that for separable programming purposes, restricted basis entry will not be required. (b) Using the same grid of points as used in Example 8.11, solve the resulting linear program. 8.15. (a) Construct a piecewise linear approximation to the separable function ƒ(x1, x2) ⫽ x 12 ⫹ x1⫺1 ⫹ 2(x2 ⫹ 2)⫺2

376


over the intervals –14 ⭐ x1 ⭐ –52 and ⫺1 ⭐ x2 ⭐ 1. Use five grid points for x1 and four grid points for x2. (b) Use the simplex method to minimize the approximating function over the region in x1 and x2. 8.16. Show that the problem Minimize

ƒ(x) ⫽ x1 ⫹ x 22 ⫹ x 33

Subject to

⫺4 ⫺6 64x⫺1 1 x2 x3 ⭐ 1

all xi ⭓ 0 can be solved using separable programming. Assuming an upper bound of 10 for all variables, formulate the separable program using five grid points for each variable. 8.17. Given the nonlinear program Minimize

ƒ(x) ⫽ x 21 ⫹ x 22 ⫹ x 23

Subject to g1(x) ⫽ 1 ⫺ x2⫺1x3 ⭓ 0 g2(x) ⫽ x1 ⫺ x3 ⭓ 0 h1(x) ⫽ x1 ⫺ x 22 ⫹ x2x3 ⫺ 4 ⫽ 0 0 ⭐ x1 ⭐ 5 0 ⭐ x2 ⭐ 3 0 ⭐ x3 ⭐ 3 Carry out one iteration of the modified SLP algorithm of PalaciosGomez with x 0 ⫽ (4, 2, 2), ␦ ⫽ 1, and ␧ ⫽ 10⫺3. 8.18. Reformulate the pooling problem discussion in Example 8.7 for solution using the separable programming strategy. Use five grid points for flows and four for the composition. Can the resulting problem be solved using a library LP routine? Are any modifications required? 8.19. (a) Construct a piecewise linear approximation to the function ƒ(x) ⫽ 10 sin x over the interval 0 ⭐ x ⭐ 2␲ rad using nine grid points. (b) What is the maximum error in your approximating function? (c) Show that if the simplex LP method is applied to minimize the approximating function, then the minimum must occur at a grid point.

PROBLEMS

377

8.20. Given the problem Minimize

ƒ(x) ⫽ x 21 ⫹ x 22 ⫹ x 23

Subject to x1x2x3 ⫽ 9 x1 ⫹ x 22 ⭐ 6 x 22 ⫹ x1 ⭐ 4

x1, x2 ⭓ 0

How would you set up the problem for solution using the separable programming method? Can the problem be solved using this method? 8.21. Consider the problem Minimize Subject to

ƒ(x) ⫽ (x1 ⫺ x2)2 ⫹ x2 x1 ⫹ x2 ⭓ 1 x1 ⫹ 2x2 ⭐ 3 x1 ⭓ 0

x2 ⭓ 0

with initial point x 0 ⫽ (3, 0). Carry out three iterations of the Frank–Wolfe algorithm.

9 DIRECTION GENERATION METHODS BASED ON LINEARIZATION The linearization-based algorithms that we examined in Chapter 8 have the common characteristic of using LP solution techniques to specify the sequence of intermediate solution points. Given a point x (t), the linearized subproblem at x (t) is updated in some prescribed fashion, and then the exact location of the next iterate x (t⫹1) is determined by the LP procedure. Yet we know that away from the base point for the linearizaton, the linearized subproblem cannot be expected to give a very good estimate of either the boundaries of the feasible region or the contours of the objective function. Rather than relying on the admittedly inaccurate linearization to define the precise location of a point, perhaps it is more realistic to utilize the linear approximations only to determine a locally good direction for search. The location of the optimum point along this direction could then be established by direct examination of values of the original objective and constraint functions along that direction rather than by recourse to the old linearized constraints, which lose their accuracy once a departure from the base point is made. This strategy is analogous to that employed in gradient-based unconstrained search methods: A linear approximation (the gradient) is used to determine a direction, but the actual function values are used to guide the search along this direction. Of course, in the constrained case the linearization will involve the objective function as well as the constraints, and the directions generated will have to be chosen so that they lead to feasible points. This is the motivation for the methods to be discussed in this chapter. 9.1

METHOD OF FEASIBLE DIRECTIONS

Chronologically, the first of the direction-generating methods based on linearization have been the family of feasible direction methods developed by Zoutendijk [1]. 378


9.1


379

Consider the inequality-constrained problem Minimize

ƒ(x)


j ⫽ 1, . . . , J

Suppose that x (1) is a starting point that satisfies all the constraints, that is, gj (x (1)) ⭓ 0, j ⫽ 1, . . . , J, and suppose that a certain subset of these constraints are binding at x (1). Zoutendijk postulated that a direction d would be a good direction for search if it were a descent direction, that is, ⵜƒ(x (1)) 䡠 d ⬍ 0

and if, for at least a small distance along the ray, x(␣) ⫽ x (1) ⫹ ␣d

with ␣ ⭓ 0

the points x along d were feasible. To a first-order approximation, the points x(␣) along d will be feasible if, for all constraints that are binding at x (1), g˜ j (x; x (1)) ⬅ gj (x (1)) ⫹ ⵜgj (x (1))(x ⫺ x (1)) ⭓ 0 Since, by assumption, gj (x (1)) ⫽ 0 and x ⫺ x (1) ⫽ ␣d

with ␣ ⭓ 0

it follows that this requirement is equivalent to requiring that the directions d satisfy ⵜgj (x (1))d ⭓ 0

for all constraints gj (x) that are binding at x (1). A direction d that satisfies the above inequalities at x (1) is appropriately called a feasible direction. Zoutendijk’s basic idea, then, is at each stage of the iteration to determine a vector d that will be both a feasible direction and a descent direction. This is accomplished numerically by finding a normalized direction vector d and a scalar parameter ␪ ⬎ 0 such that ⵜƒ(x (1))d ⭐ ⫺␪

and ⵜgj (x (1))d ⭓ ␪

380

DIRECTION GENERATION METHODS BASED ON LINEARIZATION

and ␪ is as large as possible. In carrying out the numerical solution of this direction-finding problem, the candidate d’s can be normalized conveniently by imposing the bounds ⫺1 ⭐ di ⭐ ⫹ 1, i ⫽ 1, . . . , N. The vector d that is determined in this manner will be a reasonable compromise between choosing a direction of maximum descent and following a path into the interior of the feasible region that avoids the constraints. Once the direction vector is calculated, the next intermediate point can be found by searching on ␣ along the line x ⫽ x (1) ⫹ ␣d (1) until either the optimum of ƒ(x) is reached or else some constraint is encountered, whichever occurs first. Typically one would first conduct a line search to find the value ␣ at which some constraint gj (x) first becomes binding. That is, for each constraint gj (x) we find the value ␣j ⬎ 0 at which gj (x (1) ⫹ ␣d (1)) ⫽ 0, and then we choose ␣ to be the smallest of these ␣j’s. Numerically ␣ can be determined by first using a bracketing search to find a value of x at which some gj (x) ⬍ 0, followed by a root-finding scheme such as bisection. With ␣ known, we could then use any line search to find the ␣ in the range 0 ⭐ ␣ ⭐ ␣ that would minimize ƒ(x (1) ⫹ ␣d (1)) 9.1.1

Basic Algorithm

At a given feasible point x (t), let I (t) be the set of indices of those constraints that are active at x (t) within some tolerance ␧, that is, the active constraint set is defined as I (t) ⫽ {j: 0 ⭐ gj (x (t)) ⭐ ␧, j ⫽ 1, . . . , J} for some small ␧ ⬎ 0. A complete iteration of a feasible direction method then consists of the following three steps: Step 1. Solve the linear programming problem Maximize

␪

Subject to

ⵜƒ(x (t))d ⭐ ⫺␪ ⵜgj (x (t))d ⭓ ␪ ⫺1 ⭐ di ⭐ 1

j␧I (t) i ⫽ 1, . . . , N

Label the solution d (t) and ␪ (t). Step 2. If ␪ (t) ⭐ 0, the iteration terminates, since no further improvement is possible. Otherwise, determine

9.1


381

␣ ⫽ min{␣: gj (x (t) ⫹ ␣d (t)) ⫽ 0, j ⫽ 1, . . . , J and ␣ ⭓ 0}

If no ␣ ⬎ 0 exists, set ␣ ⫽ ⬁. Step 3. Find ␣ (t) such that ƒ(x (t) ⫹ ␣ (t)d (t)) ⫽ min{ƒ(x (t) ⫹ ␣(d (t))): 0 ⭐ ␣ ⭐ ␣} Set x (t⫹1) ⫽ x (t) ⫹ ␣ (t)d (t) and continue. Note that the definition of the set It of active constraints at x (t) is given using a tolerance parameter ␧ ⬎ 0. We will discuss the importance of this parameter in the next section. Example 9.1

Minimize

ƒ(x) ⫽ (x1 ⫺ 3)2 ⫹ (x2 ⫺ 3)2

Subject to g1(x) ⫽ 2x1 ⫺ x22 ⫺ 1 ⭓ 0 g2(x) ⫽ 9 ⫺ 0.8x21 ⫺ 2x2 ⭓ 0 The gradients of the problem functions are given by ⵜƒ ⫽ [2(x1 ⫺ 3), 2(x2 ⫺ 3)] ⵜg1 ⫽ [2, ⫺2x2]

ⵜg2 ⫽ [⫺1.6x1, ⫺2]

Suppose the feasible starting point x (1) ⫽ (1, 1) is given. At this point, g1(x (1)) ⫽ 0.0

g2(x (1)) ⫽ 6.2 ⬎ 0

Thus, g1 is the only binding constraint, I (1) ⫽ {1}, and the first subproblem to be solved becomes Maximize

␪

Subject to ⫺4d1 ⫺ 4d2 ⫹ ␪ ⭐ 0 2d1 ⫺ 2d2 ⫺ ␪ ⭓ 0 ⫺1 ⭐ (d1, d2) ⭐ 1

The solution to this linear program is d (1) ⫽ (1, 0) We must now search along the ray

with ␪ (1) ⫽ 2

382


冉冊冉冊冉冊 1 1 1⫹␣ ⫹ ␣⫽ 1 0 1

x⫽

and

␣⭓0

to find the point at which the boundary of the feasible region is intersected. Since g1(␣) ⫽ 2(1 ⫹ ␣) ⫺ 1 ⫺ 1 ⫽ 2␣ g1(x) is positive for all ␣ ⭓ 0 and is not violated as ␣ is increased. To determine the point at which g2(x) will be intersected, we solve g2(␣) ⫽ 9 ⫺ 0.8(1 ⫹ ␣)2 ⫺ 2 ⫽ 0 to obtain ␣ ⫽ 1.958. Finally, we search on ␣ over the range 0 ⭐ ␣ ⭐ 1.958 to determine the optimum of ƒ(␣) ⫽ [(1 ⫹ ␣) ⫺ 3)2 ⫹ (1 ⫺ 3)2 ⫽ (␣ ⫺ 2)2 ⫹ 4 The minimum clearly occurs at the upper bound of ␣, namely, at ␣ (1) ⫽ 1.958. Thus, the iteration terminates with x (2) ⫽ (2.958, 1), at which point only g2(x) is binding. Note that in any computer code for this algorithm, the search for the boundary as well as the line search for the optimum would have been carried out using iterative methods such as those discussed in Chapter 2. The second iteration commencing at x (2) would first involve solving the subproblem Maximize

␪

Subject to ⫺0.084d1 ⫺ 4d2 ⫹ ␪ ⭐ 0 ⫺4.733d1 ⫺ 2d2 ⫺ ␪ ⭓ 0 ⫺1 ⭐ (d1, d2) ⭐ 1

constructed using the gradient g2(x). The solution obtained by solving the small linear program is d (2) ⫽ (⫺1, 0.8028)

with ␪ (2) ⫽ 3.127

The search along the ray x⫽

冉冊冉

冊

2.958 ⫺1 ⫹␣ 1 0.8028

and

␣⭓0

will yield ␣ ⫽ 0.9311, at which point constraint g1(x) is intersected. The minimum value of ƒ(x) along the ray occurs at ␣; therefore, ␣ (3) ⫽ 0.9311

9.1


383

and x (3) ⫽ (2.027, 1.748). The iterations can be continued in this fashion until the minimum point x* ⫽ (2.5, 2.0) is reached within specified accuracy. The iterations carried out above are shown in Figure 9.1. 9.1.2

Active Constraint Sets and Jamming

From Example 9.1 it is evident that the successive direction-generating subproblems of the feasible directions method differ in the active constraint set I (t) use at each iteration. Clearly, there is a computational gain in employing only a subset of all constraints to define the subproblem, since then the direction-generating LP is smaller. However, by considering only the constraints that are actually binding at the current feasible point, a zigzag iteration pattern results, which unfortunately slows down the progress of the iterations. Example 9.2

Minimize

ƒ(x) ⫽ ⫺x1

Subject to g1(x) ⫽ x2 ⭓ 0

g2(x) ⫽ 1 ⫺ x2 ⭓ 0

g3(x) ⫽ x1 ⭓ 0

g4(x) ⫽ 8 ⫺ x1 ⭓ 0

Suppose the feasible point x (1) ⫽ (1, 0) is given. Then, since g1(x) is tight, the direction-finding subproblem is


384


Maximize

␪

Subject to

(⫺1, 0)d ⭐ ⫺␪ (0, 1)d ⭓

␪

⫹1 ⭓ d ⭓ ⫺ 1

Equivalently, we have d1 ⭓ ␪

d2 ⭓ ␪

1 ⭓ (d1, d2) ⭓ ⫺1 The solution is d (1) ⫽ (1, 1), with ␪ ⫽ 1. Moreover, the search of the line x⫽

冉冊冉冊 1 1 ⫹␣ 0 1

will clearly lead to ␣ ⫽ 1. The minimization of ƒ(x) on the interval 0 ⭐ ␣ ⭐ 1 then gives x (2) ⫽ (2, 1). At x (2), constraint g2(x) is tight, and thus the direction subproblem becomes Maximize

␪

Subject to d1 ⭓ ␪

⫺d2 ⭓ ␪

⫹ 1 ⭓ (d1, d2) ⭓ ⫺ 1

The resulting direction is d (2) ⫽ (1, ⫺1). Moreover, ␣ ⫽ 1 and x (3) ⫽ (3, 0). Evidently the iterations will zigzag across the feasible region as shown in Figure 9.2 until the upper bound on x1 is reached. Obviously, if the proximity of both bounds on x2 were taken into account at each step, a flatter search direction might have been selected and the optimum found more expeditiously. In fact, the simple definition of the active constraint set used in the basic form of the feasible direction algorithm, namely,

Figure 9.2. Plot of Example 9.2.

9.1


385

I (t) ⫽ {j: 0 ⭐ gj (x (t)) ⭐ ␧, j ⫽ 1, . . . , J} can not only slow down the progress of the iterations but also lead to convergence to points that are not Kuhn–Tucker points (ref. 2, Chap. 13). Twodimensional examples of this phenomenon are not known, but a three-dimensional case is given by Wolfe [3]. This type of false convergence is known as jamming and, loosely speaking, occurs because the steps that the method takes become shorter and shorter as the direction vectors alternate between closely adjacent boundaries. The steps become shorter not because line search optima are reached but because a nearby constraint not considered in the direction generation subproblem is encountered. Various modifications have been presented that can guarantee the elimination of jamming. We will briefly consider two of these: the ␧-perturbation method [1] and the method of Topkis and Veinott [4]. The ␧-Perturbation Method. The definition of the active constraint set I (t) given above involves a small, positive tolerance parameter ␧. As shown in reference 2, the following iterative adjustment of ␧ will lead to guaranteed

convergence: 1. At iteration point x (t) and with given ␧(t) ⬎ 0, define I (t) and carry out step 1 of the basic algorithm. 2. Modify step 2 with the following: If ␪ (t) ⭓ ␧(t), set ␧(t⫹1) ⫽ ␧(t) and continue. However, if ␪ (t) ⬍ ␧(t), set ␧(t⫹1) ⫽ –12 ␧(t) and proceed with the line search of the basic method. If ␪ (t) ⬍ 0, then a Kuhn-Tucker point has been found. With this modification, it is efficient to set ␧ rather loosely initially so as to include the constraints in a larger neighborhood of the point x (t). Then, as the iterations proceed, the size of the neighborhood will be reduced only when it is found to be necessary. Topkis–Veinott Variant. An alternative approach to the problem is to simply dispense with the active constraint concept altogether and redefine the direction-finding subproblem as follows:

Maximize Subject to

␪ ⵜƒ(x (t))d ⭐ ⫺␪

gj (x (t)) ⫹ ⵜgj (x (t))d ⭓ ␪

j ⫽ 1, . . . , J

1 ⭓ d ⭓ ⫺1 The remainder of the basic algorithm is unchanged. Note that the primary difference in the subproblem definition is the inclusion of the constraint value

386


at x (t) in the inequality associated with each constraint. If the constraint is loose at x (t), that is, gj (x (t)) ⬎ 0, then the selection of d is less affected by constraint j, because the positive constraint value will counterbalance the effect of the gradient term. As shown in reference 4, this definition of the subproblem ensures that no sudden changes are introduced in the search direction as constraints are approached, and thus jamming is avoided. Example 9.3

Consider the solution of Example 9.1 using the Topkis–Veinott variant. At x (1) ⫽ (1, 1), the direction-finding subproblem becomes ␪

Maximize

⫺4d1 ⫺ 4d2 ⫹ ␪ ⭐ 0

Subject to

2d1 ⫺ 2d2 ⭓ ␪ 6.2 ⫺ 1.6d1 ⫺ 2d2 ⭓ ␪ 1 ⭓ (d1, d2) ⭓ ⫺1 Note that the third inequality corresponds to g2(x (1)) ⫹ ⵜg2(x (1))d ⭓ ␪ The direction obtained in Example 9.1, namely d (1) ⫽ (1, 0), is the solution in this case also. The positive value of g2(x (1)) is sufficiently large to cancel out the gradient term, ⫺1.6(1) ⫺ 2(0) ⫽ ⫺1.6, so that the inequality corresponding to loose constraint g2 becomes redundant. At x (2) ⫽ (2.958, 1.0), the subproblem is Maximize

␪

Subject to ⫺0.084d1 ⫺ 4d2 ⫹ ␪ ⭐ 0 ⫺4.733d1 ⫺ 2d2 ⭓ ␪

3.916 ⫹ 2d1 ⫺ 2d2 ⭓ ␪ 1 ⭓ (d1, d2) ⭓ ⫺1 The third inequality corresponds to g1, which is loose at x (2). In this case, the optimal solution is affected by the presence of the g1 inequality. The direction obtained is d (2) ⫽ (⫺0.5816, 0.4669) with ␪ (2) ⫽ 1.8189. The search of the line

9.1

x⫽


冉冊冉

387

冊

2.598 ⫺0.5816 ⫹␣ 1.0 0.4669

leads to an optimum point x (3) ⫽ (2.0269, 1.7475), which is very slightly different from the point obtained in Example 9.1. Further iterations will show that as the iterates approach the vicinity of the optimum, the inclusion of the loose constraints serves to improve the progress toward the optimum, since it leads to somewhat steeper search directions.

9.1.3

Discussion

The basic algorithm in either its ␧ form or the Topkis–Veinott form can be shown to converge to Kuhn–Tucker points under the reasonable assumptions that the problem functions are continuously differentiable [2, 4]. In the case of nonconvex feasible regions, the line search must, of course, be carried out carefully, since it is possible that the direction d, while initially feasible, may leave the feasible region and then reenter. Normally, only the closest feasible segment would be considered. Similarly, with a nonconvex objective function, multiple local minima could be encountered in the line search. Normally, the first local minimum encountered would be accepted. Under these provisions, the method has general applicability and has the important advantage of generating only feasible iteration points. That is an important consideration in engineering applications in which the system model employed may simply be invalid or meaningless outside the feasible region. These two features, guaranteed convergence for nonconvex problems and generation of feasible points, are desirable features of feasible direction methods. However, feasible direction methods have the disadvantages of slow rate of convergence and inability to satisfactorily accommodate equality constraints. From the definition of the direction generation problem, the directions generated are gradient descent directions. Moreover, as we noted in Example 9.2, the requirement that the directions be strictly feasible causes a deflection from the steepest descent direction, which can potentially retard the rate of progress toward the solution. Thus, it can be expected that the rate of convergence will be no better than that attained by the unconstrained gradient method. No procedure has yet been proposed for accelerating the slow gradient-type convergence. Such procedures are not likely to be found, because of the propensity of the method for leaving the active constraint surface and thus complicating the possibilities of accumulating information about the constraint surface for use in accelerated steps. The feasible direction machinery cannot directly handle nonlinear equality constraints, because there is no feasible interior to such constraints. The linearization of a nonlinear equality constraint will unavoidably generate a direction that leaves the constraint surface and thus will lead to infeasible points. To accommodate equalities, each constraint hk(x) ⫽ 0 must be relaxed to allow limited excursion beyond the constraint surface, that is,

388


⫺␧ ⭐ hk(x) ⭐ ␧

However, if ␧ is small, then the line search will lead to very short steps and very slow progress. On the other hand, if the feasible direction constructions are relaxed to allow limited excursion beyond the constraint surface, then iterative solution of the constraint equations will be required to return to the feasible region, as shown in Figure 9.3. A further major disadvantage of feasible direction methods is the continued necessity to solve LP subproblems. From our experience with the optimum gradient method for unconstrained problems, we recall that the locally best direction is not necessarily best from a global point of view. Thus, it might be that the effort involved in calculating the locally best compromise between descent and feasibility is not really warranted. Perhaps the computational savings achieved in selecting a descent direction without solving a linear program will more than offset the reduction in the rate of decrease achieved per iteration. The methods we discuss in the next two sections address these weaknesses of feasible direction methods. 9.2 SIMPLEX EXTENSIONS FOR LINEARLY CONSTRAINED PROBLEMS

The motivation for the methods considered in this section stems from the observation that since at a given point the number of search directions that are both descent directions and feasible directions is generally infinite, one ought to select candidate directions based on ease of computation. In the case of linear programs, the generation of search directions was simplified by changing only one variable at a time; feasibility was ensured by checking sign restrictions, and descent was ensured by selecting a variable with negative relative-cost coefficient. In this section we consider the possibilities for modifying the very convenient simplex direction generation machinery to accommodate nonlinear functions. We restrict our attention to linearly con-

Figure 9.3. Iteration to reach the equality constraint surface.

9.2

SIMPLEX EXTENSIONS FOR LINEARLY CONSTRAINED PROBLEMS

389

strained NLP problems in order to simplify the constructions. The extension to the full NLPs will be made in Section 9.3. 9.2.1

Convex Simplex Method

We begin with a brief review of the essential steps of the (linear) simplex method expressed in vector–matrix form and applied to the standard form of the problem: Minimize

ƒ(x) ⫽ cx

Subject to Ax ⫽ b

x⭓0

Given a feasible corner point x 0, the x variable vector is partitioned into two sets: the basic variables xˆ, which are all positive, and the nonbasic variables x, which are all zero. Note that if A has M rows and x has N components, then xˆ is an M vector and x an N ⫺ M vector. Corresponding to this partition of the variables, the columns of the coefficient matrix A are partitioned into B and A and the objective function coefficient vector c into cˆ and c. The first step of the algorithm then consists of expressing the basic variables in terms of the nonbasic variables. Thus, Bxˆ ⫹ Ax ⫽ b is written as xˆ ⫹ B⫺1Ax ⫽ B⫺1b

(9.1)

by reducing the coefficients of the xˆ variables to row echelon form. In this form, the values of the current basic variables will be given by xˆ ⫽ B⫺1b while the current nonbasic variables will all be equal to zero: x⫽0 The next step of the simplex method is to calculate the vector of relativecost coefficients by forming the inner product of the basic variable cost coefficients with the nonbasic variable tableau column and subtracting the result from the cost coefficient of the nonbasic variable. Equivalently, c˜ ⫽ c ⫺ cˆ B⫺1A

(9.2)

The nonbasic variable to enter is selected by finding c˜s such that c˜ s ⫽ min{˜ci: c˜ i ⬍ 0, i ⫽ 1, . . . , N ⫺ M} The basic variable xˆ r to leave the basis is selected using the minimum-ratio rule. That is, we find r such that

390


⌬r ⬅

再

冎

xˆ j xˆ r ⫽ min : p ⬎ 0, j ⫽ 1, . . . , M prs pjs js

(9.3)

where pjk are the elements of the matrix B⫺1A. The new basic feasible solution is thus xˆ (1) ⫽ j

再

⌬r

xˆ j0 ⫺ pjs⌬r

if j ⫽ r if j ⫽ r

(9.4)

and all other variables are zero. At this point, the variables xˆ and x are relabeled. Since an exchange will have occurred, B, A, cˆ, and c will be redefined. The matrix B⫺1 is recomputed, and another cycle of iterations is begun. Suppose we apply the same algorithm but replace the linear objective function coefficients c with the linearized form of a nonlinear objective function ƒ(x). Thus, instead of c we have ⵜƒ(x 0) evaluated at some initial point x 0. If x 0 is a basic feasible solution and x is again partitioned into xˆ and x, then we can calculate the relative-cost factor, ˜ 0) ⫽ ⵜƒ(x 0) ⫺ ⵜƒ(x ˆ 0)B⫺1A ⵜƒ(x

(9.5)

in a fashion analogous to the linear case [Eq. (9.2)]. Here ⵜƒˆ denotes the partial derivatives of ƒ with respect to the xˆ variables and ⵜƒ those with respect to the x variables. In this expression, ⵜƒ˜ gives the relative change of the objective function ƒ(x) at x 0, taking into account the linear constraints. This fact can be verified as follows. Since xˆ and x are related by Eq. (9.1), we can eliminate xˆ from ƒ(x) to obtain ƒ(x) ⫽ ƒ(ˆx, x) ⫽ ƒ(B⫺1b ⫺ B⫺1Ax, x) Since ƒ(x) can be expressed as a function of x alone, we can apply the chain rule to calculate the derivative of ƒ with respect to x and thus obtain Eq. (9.5). If any component of ⵜƒ˜ is negative, then, since to a first-order approximation ˜ 0)(x ⫺ x 0) ƒ(x) ⫺ ƒ(x 0) ⫽ ⵜƒ(x

(9.6)

increasing the corresponding x variable will lead to a reduction in the objective function. As in the linear case, when we continue to increase that nonbasic variable, eventually a point is reached at which a basic variable is driven to zero. However, contrary to the linear case, once we depart from the point x 0, ⵜƒ˜ will change values, and hence the possibility exists of finding a minimum of ƒ(x) before reaching the adjacent corner point. Evidently, we will need to conduct a line search to find such a minimum. The endpoint of the

9.2


391

line search will, however, be established using the minimum-ratio rule as in the purely linear case. In summary, then, the introduction of a nonlinear objective function in the simplex machinery requires: 1. Evaluation of the function gradient 2. Calculation of the modified relative-cost factor via Eq. (9.5) 3. Execution of a line search between the current point and the selected adjacent corner point Before proceeding with further developments, let us consider a small example. Example 9.4

Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ (x2 ⫺ 2)2

Subject to

2x2 ⫺ x1 ⭐ 2 x1 ⫹ x2 ⭐ 4

x⭓0

As shown in Figure 9.4, the unconstrained optimum to the problem lies at the point (1, 2), which is infeasible. The constrained optimum lies on the face of the first constraint. Suppose we rewrite the constraint set in canonical form by introducing slack variables x3 and x4. Thus,


392


⫺x1 ⫹ 2x2 ⫹ x3 ⫽ 2

x1 ⫹ x2 ⫹ x4 ⫽ 4 The initial basis will consist of x3 and x4; hence, x1 ⫽ x2 ⫽ 0, and our starting point lies at the origin of Figure 9.4. Corresponding to xˆ ⫽ (x3, x4) and x ⫽ (x1, x2), we have B⫽

冉冊 1 0

0 1

and

A⫽

冉冊 ⫺1

1

2 1

and ⵜƒ ⫽ (⫺2, ⫺4)

ⵜƒˆ ⫽ (0, 0)

Note that B⫺1 ⫽ B. The relative-cost vector calculated using Eq. (9.5) is

冉冊冉冊

ⵜƒ˜ ⫽ (⫺2, ⫺4) ⫺ (0, 0)

1 0 0 1

⫺1

1

2 ⫽ (⫺2, ⫺4) 1

The nonbasic variable to enter will be x2 ⫽ x2, since

冏冏冏冏 ⭸ƒ˜ ⭸ƒ˜ ⬍ ⭸x1 ⭸x2

The basic variable to leave will be xˆ 1 ⫽ x3, since min(–22 , –14 ) ⫽ –22 ⫽ 1 ⫽ ⌬1 From Eq. (9.4), the new point is thus xˆ (1) 1 ⫽ x2 ⫽ ⌬1 ⫽ 1 xˆ (1) 2 ⫽ x4 ⫽ 4 ⫺ (1)⌬1 ⫽ 3 A line search between x 0 and x (1) is now required to locate the minimum of ƒ(x). Note that x1 remains at 0, while x2 changes as given by x2 ⫽ 0 ⫹ ␣(1) ⫽ ␣ Therefore, ƒ(␣) ⫽ (␣ ⫺ 2)2. Clearly, ƒ(␣) is monotonically decreasing in ␣ and the optimum ␣ ⫽ 1. The search thus advances to the adjacent corner point x (1). At x (1), we have

9.2


393

xˆ ⫽ (x2, x4) ⫽ (1, 3) x ⫽ (x1, x3) ⫽ (0, 0) and B⫽

冉冊 2 1

0 1

A⫽

冉冊 ⫺1

1 0

1

B⫺1 ⫽

冉冊 –12 ⫺–12

0 1

Moreover, ˆ (1)) ⫽ (⫺2, 0) ⵜƒ(x

ⵜƒ(x (1)) ⫽ (⫺2, 0)

Using Eq. (9.5),

冉冊冉冊 –12 ⫺–12

˜ ƒ(x (1)) ⫽ (⫺2, 0) ⫺ (⫺2, 0) ⵜ

0 1

⫺1

1 ⫽ (⫺3, 1) 0

1

Clearly, x1 ⫽ x1 should be increased. Since the nonbasic variable coefficients are given by B⫺1A ⫽

冉

⫺–12 ⫹–32

冊

–12 ⫺–12

B⫺1b ⫽

冉冊 1 3

from Eq. (9.3),

再冎

⌬2 ⫽ min

3 –23

⫽2

Thus, xˆ2 ⫽ x4 leaves the basis; and the new corner point is 1 xˆ (2) 1 ⫽ x2 ⫽ 1 ⫹ – 2⌬ ⫽ 2

xˆ (2) 2 ⫽ x1 ⫽ ⌬ ⫽ 2 A line search must now be executed on the line

冢冣冢冣冦冢冣冢冣冧

x1 0 x2 1 x⫽ ⫽ ⫹␣ x3 0 x4 3 Substituting into ƒ(x), we have

2 0 2 1 ⫺ 0 0 0 3

394


ƒ(␣) ⫽ (2␣ ⫺ 1)2 ⫹ (␣ ⫺ 1)2 Setting ⭸ƒ/ ⭸␣ ⫽ 0, it follows that ␣* ⫽ 0.6. Thus, the optimum point lies at x (2) ⫽ (1.2, 1.6, 0, 1.2) Since the search stops short of the next corner point, the old basic variable x4 will not go to zero, and hence a change of basis is not required. Suppose we retain the old basis of (x2, x4). Then ⵜƒˆ ⫽ (⫺0.8, 0)

ⵜƒ ⫽ (0.4, 0)

Therefore,

冉

˜ (2)) ⫽ (0.4, 0) ⫺ (⫺0.8, 0) ⵜƒ(x

⫺–12 –32

–12 ⫺–12

冊

⫽ (0, 0.4)

Since ⵜƒ˜ ⭓ 0, it appears that x (2) is in fact the optimum solution to the problem. Note, however, that x(2) ⫽ (x1, x3) ⫽ (1.2, 0); that is, we have a nonzero nonbasic variable or, equivalently, x (2) is not a corner point. This situation never occurs in the LP case but clearly must be handled in the extension to nonlinear objective functions. This example, in addition to illustrating the main features of the extension to nonlinear objective functions, points out another situation that must be accommodated in extending the linear simplex method, namely, the occurrence of positive nonbasic variables. Evidently, positive nonbasic variables will be introduced whenever the line search yields an optimum that is not a corner point. To accommodate this case, we first of all need to modify our criterion for selecting the nonbasic variable that is to enter the basis. Then, we will need to introduce appropriate corrections in our formulas for calculating the next corner point (or target point). ˜ (t)) vector associated At a point x (t), given a choice of basis and the ⵜƒ(x with the nonbasic variables, consider the following two situations: 1. ⵜƒ˜ i is negative for some xi, where xi may be zero or positive or 2. ⵜƒ˜ i is positive for some xi ⬎ 0 From Eq. (9.6), we have, as a first-order approximation, that

冘 ⵜƒ˜ (x ⫺ x )

N⫺M

ƒ(x) ⫺ ƒ(x (t)) ⫽

k

k

(t) k

k⫽1

Now, if ⵜƒ˜ i ⬍ 0 for some xi, then by choosing (x ⫺ x(t))i ⬎ 0 and (x ⫺ x(t))k ⫽ 0 for all other variables, we have

9.2


395

ƒ(x) ⫺ ƒ (x (t)) ⬍ 0 Thus by increasing xi, the objective function decreases. Similarly, if ⵜƒ˜ i ⬎ 0 for some xi ⬎ 0, then by choosing (x ⫺ x(t))i ⬍ 0 and (x ⫺ x(t))k ⫽ 0 for all other variables, we again have ƒ(x) ⫺ ƒ(x (t)) ⬍ 0 The choice (x ⫺ x(t))i ⬍ 0 is possible only if xi(t) ⬎ 0. Conversely, if, at x (t), ⵜƒ˜ i ⭓ 0 for all xi and ⵜƒ˜ i ⫽ 0 for all xi ⬎ 0, then no improvement is possible, and a local optimum has been obtained. We thus have deduced a rule for selecting nonbasic variables as well as a necessary condition for optimality. As will be shown in Section 9.3 in a more general setting, the conditions ⵜƒ˜ ⭓ 0

(nonnegativity condition)

(9.7)

and xi(ⵜƒ˜ i) ⫽ 0

(complementary slackness condition) for i ⫽ 1, . . . , N ⫺ M (9.8)

are simply alternative ways of expressing the Kuhn–Tucker necessary conditions for a local minimum. Example 9.5

Suppose we apply conditions (9.7) and (9.8) to the point x (2) at which the calculations of Example 9.4 terminated. At x (2), we had x ⫽ (1.2, 0)

and

ⵜƒ˜ ⫽ (0, 0.4)

Clearly, (9.7) is satisfied, since (0, 0.4) ⭓ 0. Moreover, x1 ⵜƒ˜ 1 ⫽ 1.2(0) ⫽ 0 and x2 ⵜƒ˜ 2 ⫽ 0(0.4) ⫽ 0 Therefore, conditions (9.8) are satisfied and x (2) is thus a Kuhn–Tucker point. Since ƒ(x) is convex, x (2) is in fact the global constrained minimum. Violation of either condition (9.7) or (9.8) provides an obvious construction for the selection of the nonbasic variable to enter the basis. In particular, let

396


␤s ⫽ min{0, ⵜƒ˜ i: i ⫽ 1, . . . , N ⫺ M}

(9.9)

␥q ⫽ max{0, ⵜƒ˜ i xi: i ⫽ 1, . . . , N ⫺ M}

(9.10)

If 兩␤s兩 ⬎ ␥q, then xs should be increased, adjusting only the basic variables until some xˆ goes to zero. If 兩␤s兩 ⭐ ␥q, then xq should be decreased, adjusting only the basic variables. In the later case xq should be decreased until either some xˆ goes to zero or xq itself goes to zero, whichever occurs first. We are now in a position to outline the complete steps of the algorithm known as the convex simplex method (ref. 2, Chap. 8; see also [5]). Convex Simplex Algorithm Given a feasible point x 0, a partition x ⫽ (xˆ, x) of the problem variables, and a convergence tolerance ␧ ⬎ 0. ˜ (t)). Step 1. Compute ⵜƒ(x Step 2. Compute ␤s and ␥q via Eqs. (9.9) and (9.10). Step 3. If 兩␤s兩 ⭐ ␧ and ␥q ⭐ ␧, terminate. Otherwise, consider two cases: (a) If 兩␤s兩 ⬎ ␥q, determine

再

冎

xˆ j : p ⬎ 0, j ⫽ 1, . . . , M pjs js

⌬r ⫽ min

and set ⌬ ⫽ ⌬r. (b) If 兩␤s兩 ⬍ ␥q, determine

再

⌬r ⫽ min

⫺xˆ j

pjq

冎

: pjq ⬍ 0, j ⫽ 1, . . . , M

and set ⌬ ⫽ ⫺min(⌬r, xq). Step 4. Calculate the target point v(t): v(t) ˆ (t) j ⫽ x j ⫺ pjk⌬ v(t) i ⫽

再

x(t) i ⫹ ⌬ xi(t)

if i ⫽ k otherwise

where k is equal to s or q depending upon whether 3(a) or 3(b) occurs Step 5. Find ␣* such that ƒ(x (t) ⫹ ␣*(v(t) ⫺ x (t))) ⫽ min{ƒ(x (t) ⫹ ␣(v(t) ⫺ x (t))): 0 ⭐ ␣ ⭐ 1} ␣

9.2


397

Step 6. Set x (t⫹1) ⫽ x (t) ⫹ ␣*(v(t) ⫺ x (t)). If ␣* ⫽ 1 and ⌬ ⫽ ⌬r, update the basis and the basis inverse. Otherwise, retain the same basis and go to step 1. Example 9.6

[4]

ƒ(x) ⫽ x21 ⫹ x2x3 ⫹ x24

Minimize

Subject to –14 x1 ⫹ x2 ⫹ x3 ⫹ –12 x4 ⫺ x5 ⫽ 4 –14 x1 ⫹ x2 ⫺ x6 ⫽ 2

xi ⭓ 0

i ⫽ 1, . . . , 6

Given the initial point x 0 ⫽ (0, 2, 3, 0, 1, 0) at which ƒ(x 0) ⫽ 6.0. Choose xˆ ⫽ (x2, x3)

x ⫽ (x1, x4, x5, x6)

and

Thus, B⫺1 ⫽

冉冊 0 1

1 ⫺1

ⵜƒˆ ⫽ (3, 2)

A⫽

冉

–14 –14

–12

⫺1

0

0

0 ⫺1

冊

ⵜƒ ⫽ (0, 0, 0, 0)

First Iteration Step 1. ⵜƒ˜ (x 0) ⫽ (⫺–34 , ⫺1, 2, 1) Step 2. ␤2 ⫽ min(0, ⫺–43 , ⫺1, 2, 1) ⫽ ⫺1 ␥3 ⫽ max(0, 0, 0, 2, 0) ⫽ 2 Step 3. Since 兩␤2兩 ⬍ ␥3, case (b) occurs, that is, the third nonbasic variable (x5) is to be decreased. Then, since

冉冊

冉冊冉冊冉冊

p13 0 1 ⫽ B⫺1A3 ⫽ p23 1 ⫺1

⫺1

0

⫽

0

⫺1

where A3 denotes the third column of A, we have ⌬2 ⫽ min{⫺3/ ⫺1} ⫽ 3

and

⌬ ⫽ ⫺min{3, 1} ⫽ ⫺1

Thus x5 will decrease to zero before any basic variable is driven to zero.

398


Step 4. The target point will be vˆ 1 ⫽ v2 ⫽ 2 ⫺ (0)(⫺1) ⫽ 2 vˆ 2 ⫽ v3 ⫽ 3 ⫺ (⫺1)(⫺1) ⫽ 2 vˆ 3 ⫽ v5 ⫽ 1 ⫹ (⫺1) ⫽ 0

all other vi (i.e., v1, v4, v6) will remain zero. Thus, v0 ⫽ (0, 2, 2, 0, 0, 0)

Step 5. The one-parameter minimum on the line x ⫽ x 0 ⫹ ␣(v0 ⫺ x 0)

0⭐␣⭐1

occurs at ␣* ⫽ 1.0 with ƒ(␣*) ⫽ 4.0. Step 6. The next iteration point is x (1) ⫽ v0. Since no basic variable was driven to zero, a basis change or update of B⫺1 is not required. Second Iteration Step 1. ⵜƒ˜ (x (1)) ⫽ (⫺–12 , ⫺1, 2, 0) Step 2. ␤2 ⫽ min(0, ⫺–21 , ⫺1, 2, 0) ⫽ ⫺1 ␥ ⫽ max(0, 0, 0, 0, 0) ⫽ 0 Step 3. Obviously, case 3(a) occurs. The second nonbasic variable (x4) is increased.

冉冊冉冊冉冊冉冊冉冊 p12 0 ⫽ p22 1

⌬2 ⫽ min

1 ⫺1 2 –12

–21

0

⫽

0 –21

⫽4⫽⌬

The second basic variable x3 will be forced to zero. Step 4. The next target will be vˆ 1 ⫽ v2 ⫽ 2 ⫺ 0(4) ⫽ 2 vˆ 2 ⫽ v3 ⫽ 2 ⫺ –12 (4) ⫽ 0 vˆ 2 ⫽ v4 ⫽ 0 ⫹ 4 ⫽ 4

and all other vˆ 1 ⫽ 0. Thus, v(1) ⫽ (0, 2, 0, 4, 0)

9.2


399

Step 5. The one-parameter minimum occurs at ␣* ⫽ –18 , with ƒ(␣*) ⫽ 3.625. Step 6. The new point is x (2) ⫽ (0, 2, 1.75, 0.5, 0, 0) Again, the basis need not be changed, since the line search stops short of the target point. The iterations will continue in this fashion until the convergence check in step 3 is satisfied. The foregoing example illustrates one characteristic feature of the convex simplex method: the tendency for the basis to remain unchanged over a series of iterations. This occurs because the intermediate solutions are not corner points but instead lie within the feasible region (nonzero nonbasic variables). Thus, the partition into basic and nonbasic variables essentially amounts to the selection of a reference coordinate system in the nonbasic variables. When viewed in this light, the convex simplex method essentially amounts to the one-variable-at-a-time search called sectioning in the unconstrained case. It has been found empirically that convergence can be enhanced by occasional changes of basis even if none is strictly required by the computations. Two rules for basis selection have been suggested: 1. After each iteration, order the variables by decreasing magnitude and select as basic variables the M variables largest in magnitude. In this case B⫺1 must be updated after each iteration. 2. Periodically update the basis by using a random-selection rule. Say, after four or five iterations involving no basic change, randomly select a positive nonbasic variable to enter. Clear empirical evidence indicating which of the three basis selection strategies (the basic algorithm or the two above) is best is not available. Regardless of which selection strategy is used, convergence to a Kuhn–Tucker point can be proven for continuously differentiable ƒ(x) and bounded feasible regions under the nondegeneracy assumption (the basis variables are always positive) (ref. 2, Chap. 8). The rate of convergence to non-corner-point optima can be quite slow; however, the direction-finding machinery is very simple and rapid. Thus, one would expect the solution efficiency to be better than that of the Frank–Wolfe algorithm. 9.2.2

Reduced Gradient Method

From the interpretation of the convex simplex method as a sectioning procedure within the space of the nonbasic variables, we might deduce that the method could potentially be improved by changing all nonbasic variables simultaneously. In this way the change in the nonbasic variables would more closely resemble the unconstrained gradient method. However, since both

400


condition (9.7) and condition (9.8) must be satisfied at a Kuhn–Tucker point, the nonbasic variable direction vector must be defined as follows: di ⫽

再

⫺ⵜƒ˜ i ⫺xi ⵜƒ˜ i

if ⵜƒ˜ i ⭐ 0 if ⵜƒ˜ i ⭓ 0

(9.11)

where i ⫽ 1, . . . , N ⫺ M. Thus, if ⵜƒ˜ i ⬍ 0, nonbasic variable i is increased; while if ⵜƒ˜ i ⬎ 0 and xi ⬎ 0, then nonbasic variable i is decreased. This definition ensures that when di ⫽ 0 for all i, the Kuhn–Tucker conditions will be satisfied. When d the change in the nonbasic variables, is calculated using (9.11), then the change in the basic variables must be calculated using dˆ ⫽ ⫺B⫺1Ad

(9.12)

in order that the new basic variables satisfy the linear constraints. In this fashion a direction vector is defined that both satisfies the constraints and is a descent direction. However, the variable nonnegativity conditions must also be satisfied. Thus, the step in the direction d must be such that xˆ ⫽ xˆ (t) ⫹ ␣dˆ ⭓ 0 x ⫽ x(t) ⫹ ␣d ⭓ 0 where x (t) is the current feasible point. In the first case, the limiting ␣ value will be given by

再

␣1 ⫽ min

冎

⫺xˆ (t) i for all dˆ i ⬍ 0, i ⫽ 1, . . . , M dˆ i

If all dˆ i ⭓ 0, then set ␣1 ⫽ ⬁. In the second case, the limiting ␣ value will be given by

再

␣2 ⫽ min

⫺x(t) j

dj

冎

for all dj ⬍ 0, j ⫽ 1, . . . , N ⫺ M

If all dj ⭓ 0, then set ␣2 ⫽ ⬁. Let ␣max ⫽ min(␣1, ␣2) and use ␣max as the maximum step size in the line search. The overall algorithm can now be defined as follows. Reduced Gradient Algorithm Given a feasible point x 0, a partition of the problem variables x ⫽ (xˆ, x) and a convergence tolerance ␧ ⬎ 0.

9.2


401

Step 1. Compute ⵜƒ˜ (x (t)). Step 2. Compute d via (9.11) and dˆ via (9.12). If 㛳d㛳 ⬍ ␧, terminate. Otherwise, continue. Step 3. Calculate the step-size limit ␣max as defined above. Step 4. Find ␣* such that ƒ(x (t) ⫹ ␣*d) ⫽ min {ƒ(x (t) ⫹ ␣d)⬊ 0 ⭐ ␣ ⭐ ␣max} ␣

Step 5. Calculate the new point, x (t⫹1) ⫽ x (t) ⫹ ␣*d If ␣* ⫽ ␣max ⫽ ␣1, change the basis to avoid degeneracy. Otherwise, go to step 1. Example 9.7

To illustrate the steps of the reduced gradient algorithm, consider again the problem of Example 9.6 with the initial point x 0 ⫽ (0, 2, 3, 0, 1, 0). With the same choice of basic and nonbasic variables, xˆ ⫽ (x2, x3) and x ⫽ (x1, x4, x5, x6), the reduced-cost vector will be ˜ 0) ⫽ (⫺–43 , ⫺1, 2, 1) ⵜƒ(x

and B⫺1A ⫽

冉

–14

0

0

–12

冊

0 ⫺1 1

⫺1

Step 2.

d⫽

冦

⫹ ⫹ ⫺ ⫺

–34

since since since since

1 (1)(2) (0)(1)

ⵜƒ˜ 1 ⵜƒ˜ 2 ⵜƒ˜ 3 ⵜƒ˜ 4

⬍ ⬍ ⬎ ⬎

0 0 0 0

Then,

dˆ ⫽ ⫺

冉

1 4

–

0

0

0

–12

⫺1

冊冢

⫺1

1

3 5 Therefore, d ⫽ (–34 , ⫺–– 16 , ⫺– 2 , 1, ⫺2, 0).

冣

–34

1

⫺2

0

3 5 ⫽ (⫺ –– 16 , ⫺ – 2)

402


Step 3. Step-size check:

再再冎

冎

␣1 ⫽ min

⫺2 ⫺3 , 3 ⫺ (–– ) ⫺ (–52 ) 16

␣2 ⫽ min

⫺1 (⫺2)

⫽ 1.2

⫽ –12

␣max ⫽ min(1.2, –21 ) ⫽ –21

Therefore, at the target point, v0 ⫽ x 0 ⫹ ␣maxd, the third nonbasic variable (x5) will be reduced to zero. Step 4. Line search on ␣, 0 ⭐ ␣ ⭐ ␣max. It can be shown that ƒ(␣) decreases monotonically to ␣max. Thus, the optimum ␣* is equal to the bound ␣max. Step 5. The new point is 61 7 1 x (1) ⫽ x 0 ⫹ –21 d ⫽ (–83 , –– 32 , – 4, – 2 , 0)

at which point ƒ(x (1)) ⫽ 3.727 A basis change is not required since the basic variables remain positive. The iterations will continue in this fashion. Note that in this first iteration a lower objective function value was attained than was the case with the convex simplex method (ƒ(x (1)) ⫽ 4.0). The reduced gradient method was first proposed by Wolfe [6], who sug˜ gested this name because ⵜƒ˜ can be viewed as the gradient of ƒ(x) in the reduced space of the x variables. Modifications of the method to avoid jamming were presented by McCormick [7] and are incorporated in the above outline. A convergence proof, again under the nondegeneracy assumption and the requirement that ƒ(x) be continuously differentiable, is given in reference 8, Section 10.4. Numerical comparisons [9] between this algorithm and the convex simplex method show that their performance is very similar, with the convex simplex method slightly superior in the final phase of the search. However, for larger scale implementations it seems that the reduced gradient strategy is preferable because the effort involved in conducting a line search is the same regardless of the form of the direction vector. The small additional effort required to calculate dˆ thus appears justified, since, especially farther from the solution, a gradient direction ought to be more efficient than a sectioning step. As in the convex simplex method, alternate basis selection strategies can be employed. Some implementations [10] favor changing basis only when a basic variable reaches zero (or equivalently, its upper or lower bound), since this saves recomputation of B⫺1. The nonbasic variable selected to replace

9.2


403

this basic variable is the one that is largest (furthest from its bounds). Further discussion of these matters is given in reference 10. 9.2.3

Convergence Acceleration

The convex simplex method and the reduced gradient method can be viewed as, respectively, the sectioning search and the ordinary gradient method applied in a reduced variable space, and thus both can be expected to share the generally slow final convergence rates of these unconstrained methods. Given the enhanced convergence rates attained in the unconstrained case by the use of conjugate direction and quasi-Newton constructions, it is reasonable to consider the application of these techniques in the linearly constrained case. Recall that the computation of the ⵜƒ˜ vector essentially made use of the linear constraints to eliminate the xˆ variables. The independent problem variables are thus the nonbasic variables x. Once the neighborhood of the optimum is approached and the basis no longer changes, then the succeeding iterations will effectively involve only an unconstrained search over the x variables. Thus, the gradient vector ⵜƒ˜ can be used to calculate a modified search direction vector d using either the conjugate gradient formula or one of the quasi-Newton formulas, such as the BFGS method. The latter is preferable but requires storage of the quasi-Newton matrix. Whichever update formula is used, the acceleration calculations will be effective only if the basis remains unchanged. Any change of basis will alter the set of independent variables and thus will effectively alter the shape of the reduceddimensionality objective function. As a result, the usefulness of the accumulated information about the shape of the function will be lost. The only modification of the reduced gradient method required to implement the acceleration calculations occurs in step 2, the computation of the direction vector d. For the first iteration or any iteration following a basis change, the modified step will consist of the definition of the modified ‘‘gradient’’ g, g(t) i ⫽

再

ⵜƒ˜ (t) i

0

if x(t) i ⬎ 0 otherwise

or ⵜƒ˜ (t) i ⫽ 1, . . . , N ⫺ M i ⬍ 0;

and the search vector specification d(t) ⫽ ⫺g(t). For all other iterations, g is again defined as above, but d is calculated using the selected conjugate gradient or quasi-Newton formula. For instance, if the conjugate gradient method is used, then d(t) ⫽ ⫺ g(t) ⫹

㛳g(t)㛳2 (t⫺1) d 㛳g(t⫺1)㛳2

As in the unconstrained case, if exact line searches are performed, if the objective function is quadratic, and if no basis changes occur, the exact op-

404


timum will be found in N ⫺ M iterations. For general functions, continued iterations will be required until convergence is attained. However, the linear convergence rate of the unmodified algorithm can now in principle be upgraded to be superlinear. Example 9.8

Consider the solution of the problem ƒ(x) ⫽ x21 ⫺ 2x1x2 ⫹ x22 ⫹ x23

Minimize

Subject to x1 ⫹ 2x2 ⫹ x3 ⫽ 3 xi ⭓ 0

i ⫽ 1, 2, 3

starting with the point x 0 ⫽ (1, –12 , 1) and using the reduced gradient method coupled with the conjugate gradient direction formula. First Iteration. At x 0, ⵜƒ(x 0) ⫽ (1, ⫺1, 2). Suppose we select x1 as basic variable and x2, x3 as nonbasics. Then, B ⫽ (1), A ⫽ (2, 1) and B⫺1 ⫽ (1). Step 1. ⵜƒ0 ⫽ (⫺1, 2) ⫺ (1)(1)(2, 1) ⫽ (⫺3, 1) Step 2. Next, we have d⫽

冉冊 3

冉冊 3

dˆ ⫽ ⫺ (2, 1)

and

⫺1

⫺1

⫽ ⫺5

Step 3. The maximum step will be given by

冉冊冉冊

␣1 ⫽ min ␣max ⫽ min

⫺1 ⫺5

⫽

1 5

1 1 ,1 ⫽ 5 5

Step 4. The line search will yield the optimum step 1 ␣* ⫽ –– 13 ⬍ ␣max

at which value ƒ(␣*) ⫽ 0.8654. Step 5. The new point is 8 19 12 x (1) ⫽ (–– 13 , –– 26 , –– 13 )

and a basis change is not required.

冉冊

␣2 ⫽ min

⫺1 ⫺1

⫽1

9.2


405

Second Iteration 3 3 24 Step 1. ⵜƒ(x (1)) ⫽ (⫺–– 13 , –– 13 , –– 13 ) and then 3 24 3 9 27 ⵜƒ(x (1)) ⫽ (–– 13 , –– 13 ) ⫺ (⫺–– 13 )(2, 1) ⫽ (–– 13 , –– 13 )

Step 2. The search vector d is calculated using the conjugate gradient formula, d(1) ⫽ ⫺ ⵜƒ˜ (1) ⫹ ⫽⫺ ⫽

冉

冉冊 9 –– 13 27 –– 13

⫹

㛳ⵜƒ˜ (1)㛳 0 d 㛳ⵜƒ˜ 0㛳

冉冊

9 2 27 2 ((–– 3 13 ) ⫹ (–– 13 ) ) ⫺1 32 ⫹ 1 2

冊

0.74556

⫺2.5562

Then dˆ ⫽ B⫺1Ad ⫽ ⫺ (2, 1)d ⫽ 1.06509 Step 3. The step-size limit is given by ␣max ⫽ min(␣1, ␣2) ⫽ min(⬁, 0.36111) ⫽ 0.36111

Step 4. The line search will yield the exact optimum ␣* ⫽ 0.36111. Step 5. The new point is x (2) ⫽ (1.0, 1.0, 0.0) with ƒ(x (2)) ⫽ 0.0. This is the exact solution, since ⵜƒ(x (2)) ⫽ (0, 0, 0)

As expected, the solution was obtained in two iterations, since the objective function was quadratic and the problem involved two nonbasic variables. As shown by Zangwill (ref. 2, Sect. 9.1), acceleration can also be applied with the convex simplex method. However, since this strategy reduces to a sectioning procedure in the space of the nonbasic variables, it proves convenient to use the conjugate direction method. While this is in itself an interesting mating of the linear simplex method with a direct-search method, it is inherently unsatisfactory to have available the function gradient and then resort to a direct-search method with its increased number of line searches. In summary, the adaptation of simplexlike constructions to linearly constrained NLPs that we explored in this section has led to several very useful consequences. First, the partition into basic/nonbasic variables and the as-

406


sociated array manipulations have provided a means of calculating a feasible descent direction that only requires solution of a set of linear equations. Second, the modification of the relative-cost coefficient calculation to accommodate nonlinear objective functions has provided an optimality test that not only allows us to deduce whether a point satisfies the necessary conditions for a local minimum but also is constructive so that if it is violated it can be used directly to determine a direction for improvement. Finally, the directiongenerating formulation readily permits the application of quasi-Newton-type convergence acceleration constructions. In the next section, we consider the extension of these constructions and their very desirable properties to nonlinearly constrained problems.

9.3

GENERALIZED REDUCED GRADIENT METHOD

The basic reduced gradient strategy introduced in the previous section can be extended readily to solve the general NLP problem. In this section we discuss the constructions involved in considerable detail. For simplicity we initially consider the nonlinear equality-constrained form of the NLP problem, namely, Minimize

ƒ(x)

Subject to hk(x) ⫽ 0

(9.13) k ⫽ 1, . . . , K

We begin with a generalization of the implicit variable elimination procedure ˜ As part of this analysis employed in defining the reduced gradient vector ⵜƒ. we show that the optimality criterion stated in terms of this vector are equivalent to the Lagrangian conditions. Then we develop the basic form of the algorithm, followed by a discussion of the various extensions and enhancements that are incorporated in state-of-the-art implementations of this method. 9.3.1

Implicit Variable Elimination

As we observed in our discussion of direct-search methods for constrained problems, if a problem involves nonlinear equality constraints that are simple enough that they can be solved explicitly and used to eliminate variables, then it is usually expedient to do so. From a computational point of view this device reduces not only the number of variables in the problem but also the number of constraints. For instance, if the constraint hl(x) ⫽ 0 can be solved to yield xk ⫽ ␾(x1, x2, . . . , xk⫺1, . . . , xN)

9.3


407

then it is advantageous to do so because by using the above equation to substitute out the variable xk from the objective and remaining constraint functions, both the variable xk and the constraint hl(x) are eliminated from the problem. Now, in most cases the equality constraints are algebraically complex so that such explicit elimination is not possible. In those cases a reduction in problem dimensionality may still be carried out implicitly, in a manner analogous to our treatment of linear constraints in the previous section. Suppose x (1) is a point satisfying the constraints of the equality-constrained problem (9.13), and suppose we construct a linear approximation to the problem constraints at the point x (1). The result is h˜ k(x; x (1)) ⬅ hk(x (1)) ⫹ ⵜhk(x (1))(x ⫺ x (1))

k ⫽ 1, . . . , K

Suppose we wish to use these linearized equations to predict the location of another feasible point, that is, a point at which h˜ k(x; x (1)) ⫽ 0

k ⫽ 1, . . . , K

Then, since hk(x (1)) ⫽ 0, k ⫽ 1, . . . , K, that point must satisfy the system of linear equations ⵜhk(x (1))(x ⫺ x (1)) ⫽ 0

k ⫽ 1, . . . , K

(9.14)

In general, K ⬍ N; consequently, this system of equations will have more unknowns than equations and cannot be solved to yield a unique solution. However, we can solve for K of the N variables in terms of the other N ⫺ K. For instance, suppose we select the first K variables and label them xˆ (basic) and label the remaining variables x (nonbasic). Corresponding to this partition of the x’s, suppose we partition the row vectors ⵜhk into ⵜhˆ k and ⵜhk and accumulate these subvectors into two arrays J and C. The matrix J will consist of elements

J⫽

冢冣 ⵜ1hˆ ⵜ2hˆ ⯗ ⵜkhˆ

and the matrix C of the elements

C⫽

冢冣 ⵜ1h ⵜ2h ⯗ ⵜkh

408


We can now rewrite the system of Eqs. (9.14) using the matrix notation introduced above as J(ˆx ⫺ xˆ (1)) ⫹ C(x ⫺ x(1)) ⫽ 0 Assuming that the square K ⫻ K matrix J has nonzero determinant, we can solve this set of equations for the xˆ variables. Equivalently we can invert J to obtain xˆ ⫺ xˆ (1) ⫽ ⫺J⫺1C(x ⫺ x(1))

(9.15)

To a first-order approximation, all points in the vicinity of x (1) satisfying the constraints hk(x) ⫽ 0 will be given by matrix equation (9.15). Moreover, for any choice of the nonbasic variables x, the matrix equation will calculate values of the basic variables xˆ that will satisfy the linear approximations to the original constraints. In effect, the linearization has allowed us to solve the constraints for K variables even though we could not accomplish this directly with the constraints themselves. Once we have this result, we can, of course, do what we previously did in the linearly constrained case. Namely, we can use Eq. (9.15) to eliminate variables xˆ from the objective function ƒ(x). Thus, ˜ x; x) ⯝ ƒ(ˆx (1) ⫺ J⫺1C(x ⫺ x(1)), x) ƒ(ˆ

(9.16)

and we have reduced ƒ to a function involving only the N ⫺ K nonbasic variables x. Since ƒ appears to be an unconstrained function of the nonbasic variable x, we apply the usual necessary conditions for x (1) to be a local minimum of ƒ. As with any unconstrained problem, necessary conditions for x (1) to be a local minimum of ƒ are that the gradient of ƒ with respect to x be zero. ˜ ⫽ ƒ˜ (xˆ(x), x), then Using the chain rule, since ƒ(x) ⭸ƒ˜ ⭸ƒ ⭸ƒ ⭸xˆ ⫽ ⫹ 䡠 ⭸x ⭸x ⭸xˆ ⭸x

Since, from Eq. (9.15), ⭸xˆ ⫽ ⫺J⫺1C ⭸x

and if we write ⵜƒ ⫽ ⭸ƒ/ ⭸x and ⵜƒˆ ⫽ ⭸ƒ/ ⭸xˆ, then it follows that

9.3


˜ (1)) ⫽ ⵜƒ(x (1)) ⫺ ⵜƒ(x ˆ (1))J⫺1C ⵜƒ(x

409

(9.17)

The first-order necessary conditions thus become ˆ (1))J⫺1C ⫽ 0 ⵜƒ(x (1)) ⫺ ⵜƒ(x

The vector ⵜƒ˜ defined by (9.17) is called the reduced gradient (or constrained derivative) of the equality-constrained problems. This, of course, is precisely the same construction that we obtained in the development of the convex simplex method. The first-order necessary conditions for a local minimum expressed in terms of the reduced gradient reduces simply to the statement that ˜ (1)) ⫽ 0 ⵜƒ(x

It can readily be shown that the reduced gradient optimality criterion is equivalent to the Lagrangian optimality criterion. Thus, zeros of the reduced gradient are Lagrangian stationary points. To verify this, we merely need to write the Lagrangian necessary conditions for the equality-constrained problem. These are ⵜƒ(x*) ⫺ (v*)T ⵜh(x*) ⫽ 0

where ⵜh is understood to be the matrix of constraint gradients. If we now introduce the definition of independent and dependent variables x and xˆ, then the above system of equations can be written as ˆ ⵜƒ(x*) ⫺ (v*)T J ⫽ 0 ⵜƒ(x*) ⫺ (v*)T C ⫽ 0

The first of these equations can be solved for v* to yield ⫺ T ˆ v* ⫽ (ⵜƒ(x*)J )

which when substituted into the second equation yields ⫺1 ˆ ⵜƒ(x*) ⫺ ⵜƒ(x*)J C⫽0

This is precisely the reduced gradient condition that we derived earlier. Note that the Lagrange multipliers will be calculated automatically whenever the reduced gradient is evaluated. By using linear approximations to the equality constraints we have thus accomplished two things:

410


1. We have obtained a set of linear equations that allow us to estimate the values of xˆ corresponding to any perturbation of the nonbasic variables x about the point x (1) such that the resulting point is feasible. 2. We have derived a direct first-order necessary condition for the equalityconstrained problem that implicitly accounts for the equality constraints. In the next section we will show how these results of the implicit variable elimination construction can be used to develop a reduced gradient algorithm for the general nonlinear problem. The generalized reduced gradient (GRG) method is only one of a class of algorithms based on implicit variable elimination. Other prominent members of this class that we have already considered are the simplex method of linear programming, the convex simplex method, and the reduced gradient method. 9.3.2

Basic GRG Algorithm

In the previous section, we used linearizations of the constraints to express the objective function as an unconstrained function of N ⫺ K independent variables. As a consequence of this construction, the problem can be treated as simply an unconstrained optimization problem—at least for small excursions about the base point for the linearization—and presumably it can be solved using any of a variety of unconstrained gradient methods. Thus, at least in principle, the following analog to the reduced gradient algorithm suggests itself. At iteration t, suppose the feasible point x (t) is available along with the partition x ⫽ (xˆ, x) which has associated with it a constraint gradient submatrix J with nonzero determinant. Step 1. Calculate ˆ (t))J⫺1C ⵜƒ˜ ⫽ ⵜƒ(x (t)) ⫺ ⵜƒ(x

Step 2. If 㛳ⵜƒ˜ 㛳 ⭓ ␧, stop. Otherwise, set ˜ T d ⫽ (⫺ ⵜƒ)

dˆ ⫽ ⫺J⫺1Cd ˆ d)T d ⫽ (d,

Step 3. Minimize ƒ(x (t) ⫹ ␣d) with respect to the scalar parameter ␣. Let ␣ (t) be the optimizing ␣; set x (t⫹1) ⫽ x (t) ⫹ ␣ (t)d and go to step 1.

9.3


411

It is easy to verify that regardless of the degree of nonlinearity the direction d generated in this fashion is a descent direction. From a first-order Taylor’s expansion of Eq. (9.16), we have that ˜ ˜ (t)) ⫽ ⵜƒ(x ˜ (t))(x ⫺ x(t)) ƒ(x) ⫺ ƒ(x (t)) ⬇ ƒ(x) ⫺ ƒ(x ˜ (t))d ⫽ ␣ ⵜƒ(x

Thus, if d is chosen as ˜ (t)))T d ⫽ ⫺ (ⵜƒ(x

then it follows that ˜ ⫺ ⵜƒ) ˜ T ⫽ ⫺␣㛳ⵜƒ˜ 㛳2 ƒ(x) ⫺ ƒ(x (t)) ⯝ ␣(ⵜƒ)(

The right-hand side of this expression is less than zero for all positive values of ␣. Consequently, for all positive ␣’s small enough that the linear approximations used are sufficiently accurate, it follows that ƒ(x) ⫺ ƒ(x (t)) ⬍ 0 and consequently d is a descent direction. Note that the specification dˆ ⫽ ⫺J⫺1Cd is implicit in the above construction, since Eq. (9.16) is based on using the constraint linearizations to eliminate the xˆ variables. While it is thus evident that step 2 produces a descent direction, it is not at all clear that the points generated in following that direction will be feasible. In fact, for nonlinear constraints, d will not be a feasible direction. Rather, because it is constructed using a linearization of the equality constraints, it is almost certain to lead to points away from the constraints. As can be seen from the next example, the linearization of an equality constraint at a feasible point is equivalent to the construction of a plane tangent to the constraint at the point of linearization. In general, virtually all points on the tangent plane will not satisfy the original equality. Example 9.9

Consider the linearization of the constraint h(x) ⫽ (x1 ⫺ 1)3 ⫺ x2 ⫹ 1 ⫽ 0 at the point x 0 ⫽ (–12 , –78 ). The linear approximation is

412


˜ h(x; x 0) ⫽ 0 ⫹ –34 (x1 ⫺ –12 ) ⫺ (x2 ⫺ –78 )

If we insist that all points x satisfy this linearization, that is, h˜ (x; x 0) ⫽ 0, we obtain the linear equality –34 x1 ⫺ x2 ⫹ 0.5 ⫽ 0

From Figure 9.5 it is evident that no point on this plane, with the exception of the point (2, 2), will satisfy h(x) ⫽ 0. Summarizing the preceding discussion, we have demonstrated first that d is a descent direction and second that it will in general lead to infeasible points. More precisely, we have shown that the vector d is a descent direction in the space of the nonbasic variables x but that the composite direction vector d⫽

冉冊 dˆ d

where dˆ is calculated via the linear equation dˆ ⫽ ⫺J⫺1Cd yields infeasible points. Since d has desirable properties but the addition of dˆ causes undesirable consequences, it seems reasonable that the computation of dˆ should be revised. The question is how? We can gain further insight into the situation by considering a graphical example. Consider the spherical constraint surface

Figure 9.5. Linearization of a nonlinear constraint.

9.3


413

h(x1, x2, x3) ⫽ x12 ⫹ x22 ⫹ x32 ⫺ 1 ⫽ 0 shown in Figure 9.6. Given the feasible point x 0, the linearization of h(x) ⫽ 0 at x 0 will yield a tangent plane. Suppose that x1 and x2 are chosen as nonbasic variables, that is, x ⫽ (x1, x2), and x3 as a basic variable, that is, xˆ ⫽ (x3). Furthermore, suppose that for some objective function step 2 of our prototype algorithm generates the direction d shown in the figure. The subvector d will consist of the projection of d on the x1x2 plane, while dˆ will simply be the projection of d on the x3 axis. Clearly the point v(2) that would be attained by following along d from x 0 is not feasible, that is, h(v(2)) ⫽ 0. The point v(2) is infeasible, because for the chosen change in the independent variables, x(2) ⫺ x 0, the change in the dependent variable, xˆ (2) ⫺ xˆ 0, calculated using the approximation xˆ (2) ⫺ xˆ 0 ⫽ ⫺J⫺1C(x(2) ⫺ x 0) is not accurate enough to yield a value of xˆ that satisfies the equation h(x(2), xˆ ) ⫽ 0 exactly. We should have expected this, because, as we observed previously, while the linearizations are good enough to determine local descent directions, they are not good enough to actually calculate feasible minimum points. What we really ought to do is to calculate d using the linearization, but then, rather than calculating dˆ as in step 2 and minimizing ƒ along the line fixed by d, we should project d onto the constraint surface and minimize along the re-

Figure 9.6. The reduced gradient direction vector and curve of feasible points.

414


sulting curve. Thus, in effect we will search for the minimum of ƒ(x) along an implicitly defined curve that is given by the set of ␣ and xˆ values that satisfy the equation h(x 0 ⫹ ␣ d, xˆ ) ⫽ 0 This is the curve passing through points x 0, x(1), and x(2) in Figure 9.6. If d⫽

冉冊 1 1

x 0 ⫽ (0.1, 0.1)

and

then this curve will be given by all values of ␣ and x3 satisfying (0.1 ⫹ ␣)2 ⫹ (0.1 ⫹ ␣)2 ⫹ x23 ⫺ 1 ⫽ 0 In the general case with multiple constraints, this means that for every value of ␣ that is selected as a trial the constraint equation will have to be solved for the values of the dependent variables xˆ that will cause the resulting point to be feasible. Thus step 3 of our prototype algorithm must be replaced by an iterative procedure that will, for a given value of ␣, iterate on the constraints to calculate a value of xˆ that satisfies the constraints. This can be accomplished by using, for instance, Newton’s method to solve the set of equations hk(x(t) ⫹ ␣d, xˆ ) ⫽ 0

k ⫽ 1, . . . , K

Recall from Chapter 3 that the iteration formula for Newton’s method for solving a system of K equations in K unknowns zk, hk(z) ⫽ 0

k ⫽ 1, . . . , K

involves the gradients ⵜzhk of the functions evaluated at the current best estimate z (i). If we denote the matrix with rows ⵜzhk(z (i)) by ⵜh(i) and the vector of function values hk(z (i)) by h(i), then the iteration formula will be given by z(i⫹1) ⫽ z(i) ⫺ (ⵜh(i))⫺1h(i) In the case of our constraint iterations, we have z(i) ⫽ xˆ (i) ⵜh(i) ⫽ J(x (t) ⫹ ␣d, xˆ (i))

and

9.3


415

h(i) ⫽ h(x (t) ⫹ ␣d, xˆ (i)) Therefore, the Newton iteration formula will be xˆ (i⫹1) ⫽ xˆ (i) ⫺ [J⫺1(x (t) ⫹ ␣ d, xˆ (i))] 䡠 h(x(t) ⫹ ␣ d, xˆ (i))

(9.18)

where the notation J⫺1(x(t) ⫹ ␣d, xˆ (i)) indicates that the elements of the inverse of J are functions of the step size ␣ and of the basic variables xˆ. Assuming that Newton’s method will converge to an xˆ that satisfies the constraints, then ƒ(x) can be evaluated at that point to determine if an improvement over x (t) has been achieved. If no improvement has been achieved, then ␣ must be reduced using some search logic, and the Newton iterations must be repeated. If ƒ(x) is improved, then one can either continue with a new value of ␣ until no further improvement in ƒ(x) can be obtained or else, as is more typically done, the current point can be retained and a new direction vector calculated. It is, of course, possible that for the ␣ selected the Newton iterations may fail to find an xˆ that satisfies the constraints. This indicates that ␣ was chosen too large and should be reduced. In Figure 9.6 there is no value of xˆ (⫽x3) corresponding to the point x(3) that will intersect the constraint. If this occurs, then ␣ must be reduced and another cycle of Newton iterations initiated. With these additional steps, a composite GRG algorithm can be formulated. Generalized Reduced Gradient Algorithm Given an initial feasible point x 0, a specified initial value of the search parameter ␣ ⫽ ␣0, termination parameters ␧1, ␧2, and ␧3, all ⬎ 0, and a reduction parameter ␥, 0 ⬍ ␥ ⬍ 1. Step 1. Choose a partition of x into xˆ and x such that J has nonzero determinant and calculate the reduced gradient: ˆ (t)) ⵜƒ(x

Step 2. If 㛳ⵜƒ˜ 㛳 ⭐ ␧1, stop. Otherwise, set ˜ T d ⫽ ⫺ (ⵜƒ)

dˆ ⫽ ⫺ J⫺1Cd

ˆ d)T d ⫽ (d,

Step 3. Set ␣ ⫽ ␣0. For i ⫽ 1, 2, . . . : (a) Calculate v(i) ⫽ x (t) ⫹ ␣d. If 兩hk(v(i))兩 ⭐ ␧2, k ⫽ 1, . . . , K, go to (d). Otherwise, continue. (b) Let vˆ (i⫹1) ⫽ vˆ (i) ⫺ J⫺1(v(i)) 䡠 h(v(i)) v(i⫹1) ⫽ v(i)

416


(c) If 㛳vˆ (i⫹1) ⫺ vˆ (i)㛳 ⬎ ␧3, go to (b). Otherwise, if 兩hk(v(i))兩 ⭐ ␧2, k ⫽ 1, . . . , K, go to step (d). Otherwise, set ␣ ⫽ ␥␣ and go to (a). (d) If ƒ(x (t)) ⭐ ƒ(v(i)), set ␣ ⫽ ␥␣ and go to (a). Otherwise, set x (t⫹1) ⫽ v(i) and go to step 1. We illustrate the application of this algorithm with the following example. Example 9.10

ƒ(x) ⫽ 4x1 ⫺ x22 ⫹ x23 ⫺ 12

Minimize

Subject to h1(x) ⫽ 20 ⫺ x12 ⫺ x22 ⫽ 0 h2(x) ⫽ x1 ⫹ x3 ⫺ 7 ⫽ 0 Suppose we are given the feasible starting point x (1) ⫽ (2, 4, 5), and suppose we choose x1 as nonbasic variable and x2 and x3 as basic variables. Thus, xˆ ⫽ (x2, x3)

x ⫽ (x1)

and

Note that the number of nonbasic variables is equal to the total number of variables minus the number of equality constraints. The required function derivatives are ⵜƒ ⫽ (4, ⫺2x2, 3x3), ⵜh1 ⫽ (⫺2x1, ⫺ 2x2, 0), and ⵜh2 ⫽ (1, 0, 1). Step 1. Since ⵜƒ ⫽ (4, ⫺8, 10)

ⵜh1 ⫽ (⫺4, ⫺8, 0)

ⵜh2 ⫽ (1, 0, 1)

it follows that J⫽

冉冊 ⫺8

0

0 1

C⫽

冉冊 ⫺4

ⵜƒ ⫽ (⫺8, 10)

1

Next, we calculate J⫺1 ⫽

冉冊 ⫺–18

0

0 1

冉冊冉冊

ⵜƒ˜ ⫽ (4) ⫺ (⫺8, 10)

⫺–18

0

0 1

⫺4

1

⫽ ⫺2

ⵜƒˆ ⫽ (4)

9.3


417

Step 2. The direction vector becomes d ⫽ ⫺ ⵜƒ˜ ⫽ 2 dˆ ⫽ ⫺

冉冊冉冊冉冊 ⫺–18

0 1

0

⫺4

1

(2) ⫽

⫺1 ⫺2

Step 3. Set ␣0 ⫽ 1.

冢冣冢冣冢冣

2 2 4 (a) v(1) ⫽ 4 ⫹ ␣ ⫺1 ⫽ 3 5 ⫺2 3 h1(v(1)) ⫽ ⫺5 and h2(v(1)) ⫽ 0.0 Note that the linear constraint h2(x) is satisfied. This is to be expected, since a linear approximation to a linear constraint will be the constraint itself. Thus, linear constraints will always be satisfied automatically in this algorithm. Since h1 is not satisfied, Newton’s method must be applied to the system of equations h1(4, x2, x3) ⫽ 0

h2(4, x2, x3) ⫽ 0

to solve for x2 and x3 starting with the initial estimate x2 ⫽ x3 ⫽ 3. (2)

(b) vˆ

冉冊冉冊冉冊冉冊冉冊冉冊冉冊

⫺6 0 3 ⫽ ⫺ 3 0 1 ⫽

⫺1

⫺5

0

⫽

1 1 3 0 ⫹ 3 6 0 ⫺6

⫺5

0

2.167 3

(c) Since vˆ (2) ⫽ vˆ (1), another Newton iteration must be carried out. (b) vˆ (3) ⫽

冉冊冉 13 –– 6

3

⫺

13 ⫺ –– 3

0

冊冉冊冉冊

0 1

⫺1

⫺0.69

0.0

⫽

2.008 3

If we accept this result as having converged, then we must test the constraint values [step (c)]. h1(v(3)) ⫽ ⫺0.032

and

h2(v(3)) ⫽ 0.0

Assuming that 0.032 is close enough to zero, we continue step (d). (d) v(3) ⫽ (4, 2.008, 3) ƒ(x (1)) ⫽ 5 ⬍ ƒ(v(3)) ⫽ 9

418


No improvement in ƒ(x) is achieved. The initial ␣ chosen was too large and must be reduced. Suppose we reduce ␣ to 0.4.

冢冣冢冣冢冣

2 2 2.8 (a) vˆ (4) ⫽ 4 ⫹ 0.4 ⫺1 ⫽ 3.6 5 ⫺2 4.2 (4) h1(v ) ⫽ ⫺0.80 and h2(v(4)) ⫽ 0.0 The point is infeasible and Newton iteration is required. (b) vˆ (5) ⫽

冉冊冉

3.6 ⫺7.2 ⫺ 4.2 0

冊冉冊冉冊 ⫺1

0 1

⫺0.8

3.49 4.2

⫽

0

(c) If we accept this result as having converged, then we must test h1(v(5)) ⫽ ⫺0.02

h2(v(5)) ⫽ 0.0

and

Assuming that this is close enough to zero, we proceed to test if the objective function has improved. (d) ƒ(x (1)) ⫽ 5 ⬎ ƒ(v(5)) ⫽ 4.66 Since improvement has been made, x (2) is set equal to v(5). The next iteration is begun with x (2) ⫽ (2.8, 3.49, 4.2) The solution to the problem is x* ⫽ (2.5, 3.71, 4.5). Note that at that point the reduced gradient is zero:

冉

ⵜƒ˜ ⫽ (4) ⫺ (⫺7.42, 9.0)

⫺7.42

0

冊冉冊

0 1

⫺1

⫺5

1

⫽ 4 ⫺ 4 ⫽ 0.0

As can be noted from the above example, most of the effort per iteration is involved in performing the Newton iterations. To keep these iterations to a low level, the one-parameter optimizations on ␣ are normally not carried out very exactly. In fact, in our simplified version of the GRG algorithm the first feasible improved point is accepted without any further refinement. An alternative approach is to use an ‘‘inside-out’’ strategy that involves a bracketing search followed by a quadratic interpolation step. Thus, rather than choosing a larger initial ␣ and then retracting the search if improvement is not achieved, a bracketing search is first carried out on ␣ to bound the minimum. The bounding search is initiated with a small ␣ so that the initial trials remain close to the linearization point x (t), and consequently the Newton correction iterations are less likely to encounter difficulties. As ␣ is increased during bracketing, the previously calculated Newton result can always be used as a

9.3


419

good estimate with which to start the Newton corrections at the next ␣ value. Once a bracket is established, then a quadratic (or cubic) interpolation step is carried out to obtain an improved estimate of the line search optimum ␣*. The resulting point may require further correction via Newton iterations but is not improved by further successive quadratic interpolations. The extension of the reduced gradient method to nonlinearly constrained problems is generally credited to Abadie and Carpentier [11]. However, elements of the implicit variable elimination strategy with nonlinear equalities appears in the early analytical-graphical optimal design work of Johnson as early as 1956 (see the review in ref. 12). The implicit variable elimination strategy also underlies the differential algorithm of Beightler and Wilde [13] and the constrained derivative conditions of Wilde [14] and Reklaitis and Wilde [15]. The work of Abadie and Carpentier [11] did, however, lead to the first general GRG code. 9.3.3

Extensions of Basic Method

The preceding development of the basic GRG algorithm has been restricted to equality constraints and has resulted in essentially a gradient method operating within the reduced space of the nonbasic variables. In this section we consider extensions of the basic algorithm to accommodate the general NLP problem with both upper and lower variable bounds as well as the inequality constraints Minimize

ƒ(x) aj ⭐ gj (x) ⭐ bj

Subject to

hk(x) ⫽ 0 x

(L) i

⭐ xi ⭐x

j ⫽ 1, . . . , J k ⫽ 1, . . . , k

(U) i

i ⫽ 1, . . . , N

In addition, we discuss the use of the conjugate direction or quasi-Newton formulas to modify the direction vector d generated by the basic GRG algorithm so as to attain improved terminal rates of convergence. Treatment of Bounds. Variable upper and lower bounds can be accommodated either explicitly by treating each bound as an inequality constraint or implicitly by accounting for the bounds within the appropriate steps of the algorithm. The latter approach is clearly preferable, since it results in a smaller matrix J that must be inverted. To incorporate such constraints into the algorithm in an implicit fashion, three changes are necessary:

1. A check must be made to ensure that only variables that are not on or very near their bounds are labeled as basic variables. This check is

420


necessary to ensure that some free adjustment of the basic variables can always be undertaken. This can be accomplished by simply ordering the variables according to distance from their nearest bound. Thus, at the current feasible point x (t), let U) (t) (t) (L) z(t) i ⫽ min{(xi ⫺ x i ), (x i ⫺ x i )}

(9.19)

The quantities z(t) can then be ordered by decreasing magnitude, and the i variables associated with the first K can be selected as dependent variables. Of course, this partition of the variables must also result in a nonsingular J. Therefore, one or more of the first K variables may have to be rejected in favor of variables farther down the order before a nonsingular J results. 2. The direction vector d is modified to ensure that the bounds on the independent variables will not be violated if movement is undertaken in the d direction. This is accomplished by setting

冦

0 di ⫽ 0

⫺ (ⵜƒ)i

if xi ⫽ xi(U) and (ⵜƒ)i ⬍ 0, if xi ⫽ xi(L) and (ⵜƒ)i ⬎ 0 otherwise

This modification will also ensure that, if di ⫽ 0 for all i ⫽ 1, . . . , N ⫺ K, then the Kuhn–Tucker conditions will be satisfied at that point. This fact can be verified easily by applying the Kuhn–Tucker conditions to the problem with equality constraints and two types of inequalities: g⬘ ⬅ x (U) ⫺ x ⭓ 0

gⴖ ⬅ x ⫺ x (L) ⭓ 0

Verification is left as an exercise for the reader. 3. Checks must be inserted in step 3 of the basic GRG algorithm to ensure that the bounds are not exceeded either during the search on ␣ or during the Newton iterations. As shown in Figure 9.7, this can be accomplished by simply checking if the result of step 3 satisfies the variable bounds. If at least one bound is violated, a linear interpolation can be made between x (t) and v(i) to estimate the point on this line segment at which a basic variable bound is first violated. Step 3 is then reinitiated at this point to obtain a v(i) at which xˆ (L) ⭐ vˆ (i) ⭐ xˆ (U). Treatment of Inequalities. General inequality constraints can be handled within the GRG framework either by explicitly writing these constraints as equalities using slack variables or by implicitly using the concept of active

9.3


421

Figure 9.7. Linear interpolation to satisfy variable bounds.

constraint set as in feasible direction methods. In the former case, inequalities of the form aj ⭐ gj (x) ⭐ bj are converted to equalities hK⫹j(x) ⫽ gj (x) ⫺ xN⫹j ⫽ 0 by the introduction of slack variables xN⫹j, together with the bounds aj ⭐ xN⫹j ⭐ bj This approach is attractive because it allows all constraints to be treated in a uniform fashion without further restructuring of the basic algorithm. The disadvantage is that J must include a row for each gj (x) even if the inequalities are not binding. In addition, the problem dimensionality is increased. If an active constraint strategy is employed, then at each iteration point x (t) the active inequalities must be identified and their linearizations added to

422


those of the equalities to define J and C. The logic of step 3 of the basic algorithm must be modified to include tests for the violation of the previously inactive constraint. If at target point v(i) some previously inactive constraint is violated, then the step size ␣ must be adjusted and the Newton calculations repeated until a point satisfying both the active and inactive constraints is found. At the resulting points, any tight but previously inactive constraints must be added to the active constraint set, and any active constraint now not binding must be deleted from the active set. These considerations will clearly complicate the step 3 computations and will result in the need to continually readjust the size of J and C and update J⫺1. The two approaches to the treatment of inequalities are illustrated in the next example. Example 9.11


ƒ(x) ⫽ (x1 ⫺ 3)2 ⫹ (x2 ⫺ 3)2

Subject to g1(x) ⫽ 2x1 ⫺ x22 ⫺ 1 ⭓ 0 g2(x) ⫽ 9 ⫺ 0.8x21 ⫺ 2x2 ⭓ 0 4 ⭓ (x1, x2) ⭓ 0 with initial feasible point x 0 ⫽ (1, 1)T. If the slack variable approach is used, then the constraints will be reformulated to h1(x) ⫽ g1(x) ⫺ x3 ⫽ 0 h2(x) ⫽ g2(x) ⫺ x4 ⫽ 0 The expanding dimensionality starting point becomes x 0 ⫽ (1, 1, 0, 6.2) At x 0, ⵜh1(x 0) ⫽ (2, ⫺2, ⫺1, 0) ⵜh2(x 0) ⫽ (⫺1.6, ⫺2, 0, ⫺1)

and

9.3


423

ⵜƒ(x 0) ⫽ (⫺4, ⫺4, 0, 0)

Suppose we select xˆ ⫽ (x1, x2) and x ⫽ (x3, x4). Then J⫽

冉

Step 1. ⵜƒ˜ ⫽ (0, 0) ⫺ ⫽

2

⫺1.6

冊冉

冉冊冉

⫺2 ⫺2

⫺1

C⫽

(⫺4, ⫺4) ⫺2 2 ⫺1.6 ⫺2 7.2

1 (⫺1.6, 16) 7.2

0

0 ⫺1

⫺1

冊冊

0 0 ⫺1

Step 2. Set d1 ⫽

1.6 7.2

since ⵜƒ˜ 1 ⬍ 0 and x1 is at lower bound. Set d2 ⫽

⫺1.6

7.2

since ⵜƒ˜ 2 ⬎ 0 and x2 is not on its bounds. Then dˆ ⫽ ⫺J⫺1Cd ⫽

冉冊 0.6790 0.5679

Step 3. As is evident from Figure 9.8, a larger initial step can easily be taken in this case. Suppose ␣ (0) ⫽ 2 is employed. Then v0 ⫽ x 0 ⫹ 2d ⫽ (2.3580, 2.1358, 0.4444, 1.7555)

(a) h1(v0) ⫽ ⫺1.2900 and h2(v0) ⫽ ⫺1.4752 (b) vˆ (1) ⫽ vˆ 0 ⫺ J⫺1h 1 ⫺2 ⫺1.2900 2.3580 2 ⫽ ⫺ 2.1358 ⫺ 1.6 ⫺2 ⫺1.4752 7.2 T ⫽ (2.3066, 1.4393)

冉冊

冉

冊冉

冊

Steps 3b and 3c are repeated until the point vˆ ⫽ (2.1859, 1.7110)T is reached, at which h1 ⫽ ⫺1 ⫻ 10⫺4

and

h2 ⫽ ⫺3 ⫻ 10⫺5

424


Figure 9.8. Slack variable strategy, Example 9.11.

(d) At this point, ƒ(ˆv) ⫽ 2.324 ⬍ ƒ(x 0) ⫽ 8.0 Therefore, x (1) ⫽ v ⫽ (2.1859, 1.7110, 0.4444, 1.7555)T The iterations continue in this fashion until the optimum point x* ⫽ (2.5, 2, 0, 0)T is reached. Alternatively, if the active constraint strategy is used, then the constraints are first tested at x 0 to determine which are active. Since g1(x 0) ⫽ 0.0 and g1(x 0) ⫽ 6.2, constraint 1 is active and g2 will be temporarily disregarded. If the partition xˆ ⫽ (x2), x ⫽ (x1) is selected, then J ⫽ (⫺2) ⵜƒˆ ⫽ (⫺4)

C ⫽ (2) ⵜƒ ⫽ (⫺4)

Step 1. ⵜƒ˜ ⫽ (⫺4) ⫺ (⫺4)(⫺ –12 )(2) ⫽ ⫺8 Step 2. Since x2 is away from its bounds, we can set

9.3


425

d ⫽ ⫺(⫺8) ⫽ 8 dˆ ⫽ ⫺J⫺1Cd ⫽ ⫺ (⫺ –12 )(2)(8) ⫽ ⫹8 Step 3. v0 ⫽ x 0 ⫹ ␣d ⫽

冉冊冉冊 1 8 ⫹␣ 1 8

Suppose we use ␣0 ⫽ –14 ; then v0 ⫽ (3, 3) and g1(v0) ⫽ ⫺4. Step 3b. vˆ (1) ⫽ 3 ⫺ (⫺ –16 )(⫺4) ⫽ 2.333 Steps 3b and 3c will continue until the point vˆ (3) ⫽ 2.2361 is reached, at which g1 ⯝ ⫺1 ⫻ 10⫺4. At the point, v(3) ⫽ (3, 2.2361), the objective function value is 0.5835 ⬍ ƒ(x 0). However, g2(v(3)) ⫽ ⫺2.672, so that this constraint has become violated, as shown in Figure 9.9. The step size ␣ must therefore be reduced and step 3 repeated. For instance, if ␣ (1) ⫽ –16 , we obtain v(4) ⫽ (2.3333, 2.3333)

Figure 9.9. Active constraint strategy, Example 9.11.

426


A repeat of the Newton iterations will yield the new feasible point x (1) ⫽ (2.3333, 1.9149), at which ƒ(x (1)) ⫽ 1.622

g2(x (1)) ⫽ 0.815

and

In part because of the iterations required to correct the infeasibilities that can occur among the previously inactive constraints, the active constraint strategy is not used with the GRG implementations reported in the literature [11, 16, 17]. Modification of the Direction Vector. Convergence proofs are not available

for versions of GRG as it is typically implemented. Convergence has been proved [18] for an idealized version of GRG in which: 1. Exact searches on ␣ are performed 2. An ␧-perturbation procedure is used to avoid jamming. 3. A special basis-updating procedure is introduced that may require basis changes at each iteration. 4. It is assumed that at each iteration a nondegenerate, nonsingular basis can always be found. Under these conditions it can be shown that the GRG calculations will converge to a Kuhn–Tucker point. An interesting terminal convergence rate result developed for the basic equality-constrained problem by Luenberger [19] indicates that the GRG algorithm can be expected to exhibit a performance analogous to the unconstrained gradient method. Specifically, it is shown that if {x (t)} is a sequence of iterates generated using the same variable partition and converging to x*, then the sequence of objective function values {ƒ(x (t))} converges to ƒ(x*) linearly with a ratio bounded above by [(B ⫺ b)/(B ⫹ b)]2. Quantities b and B are, respectively, the smallest and largest eigenvalues of the matrix,

冉

Q ⫽ (⫺J⫺1CT, I)HL

冊

⫺J⫺1C

I

and matrix HL is the matrix of second derivatives of the Lagrangian function L ⫽ ƒ ⫺ vTh. This result indicates that the convergence rate is strongly influenced by the choice of basis (as reflected by J⫺1C) as well as by the second derivatives of all problem functions. Note that this definition of convergence rate differs from the normal definition of the rate in terms of the error in x (t), that is, 㛳x (t) ⫺ x*㛳. However, the ratio bound on the objective function values is similar to results available for the ordinary unconstrained gradient method. Luenberger’s analysis thus clearly indicates the desirability of using convergence acceleration techniques.

9.3


427

As in the reduced gradient method for the linearly constrained case, the direction vector d can be updated using conjugate gradient or quasi-Newton methods. Provided that the variable partition is not altered, the use of such updates will substantially enhance the convergence rate. Of course, whenever a change is made in the partition, the updating formula must be restarted, since the nature of the search subspace is altered. Thus, the modification of direction vectors is most effective in the final stages of the iterations when no further significant partitioning changes are encountered. Different authors have, for one reason or another, selected different update methods. For instance, Lasdon et al. [16] recommend the use of the BFGS formula and resort to conjugate gradient methods only if storage becomes limiting, while Abadie and Carpentier [11] use the Fletcher–Reeves conjugate gradient method. In view of the generally superior performance of the BFGS update, it appears that it ought to be the preferred choice. 9.3.4

Computational Considerations

From the preceding discussion it is evident that implementation of an efficient GRG algorithm involves consideration of a variety of numerical issues: Newton iterations, linear equation solving, line searching, step-size adjustment, and so on. In this section we briefly consider the following computational enhancements to the GRG algorithm: 1. 2. 3. 4. 5.

Use of approximate Newton iterations Indirect calculation of J⫺1 Special treatment of linear problem elements Use of numerical derivatives Generation of feasible starting points

Approximate Newton Calculations. The main computational burden associated with the GRG algorithm arises from the Newton iterations. Various rules and approximation devices have been proposed to reduce this overhead. These are:

1. Using the inverse of J evaluated at x (t) without subsequent updating 2. Using the approximation suggested by Abadie and Carpentier [11], J(v(k))⫺1 ⫽ 2J(x (t))⫺1 ⫺ J(x (t))⫺1J(v(k))J(x (t)) 3. Using a quasi-Newton equation-solving method to update J⫺1 without recomputing derivatives The first approach clearly is the simplest but can lead to substantially reduced rates of convergence. The second approach requires the recomputation of the

428


constraint gradients at the intermediate iteration points v(k) but avoids recomputation of the inverse of J. The third approach is attractive, since J⫺1(x (k)) can be used as the initial value, and subsequent updates only require the constraint function values. Avoidance of Matrix Inverses. Note that although all the key formulas in terms of which the GRG algorithm is defined are expressed in terms of J⫺1, the inverse need not be explicitly calculated. For instance, to calculate the reduced gradient, it is not necessary to use ⵜƒ˜ ⫽ ⵜƒ ⫺ ⵜƒˆ J⫺1C

Instead, we first solve ⵜƒˆ ⫺ vTJ ⫽ 0 to obtain the multipliers v and then evaluate the reduced gradient via ⵜƒ˜ ⫽ ⵜƒ ⫺ vTC

Similarly, to calculate dˆ ⫽ ⫺J⫺1Cd, we first calculate Cd and then solve the linear equation set Jdˆ ⫽ ⫺Cd for dˆ . Finally, the Newton calculations would be executed by solving the linear system J ⌬xˆ ⫽ ⫺h All these matrix calculations can be readily modified to take account of any special problem structure. In particular, since large problems normally tend to be sparse—that is, the matrix (J, C) tends to have relatively few nonzero elements—specialized techniques developed for the solution of large, sparse linear equation sets can be readily adopted [17]. Thus, the GRG method appears to be particularly suitable for the solution of large-scale nonlinear problems. Treatment of Linear Elements. Further significant computational savings can be achieved if, in implementing a GRG algorithm, special account is taken of linear elements in the problem formulations. Such savings are particularly significant in the case of large problems. As proposed by Murtagh and Saunders [10], each problem function can be subdivided into linear and nonlinear components,

h(x) ⫽ hN(xN) ⫹ akxL ⫹ bk where the variables xL are those that do not appear in any nonlinear terms, while the variables xN are those that appear in at least one nonlinear term. The advantages of this classification are twofold. First, the columns of the matrix of constraint gradients corresponding to the xL variables will always remain constant and hence need not be recomputed. Second, the classification

9.3


429

leads to an effective reduction in the dimensionality of the space of nonbasic variables. Following Murtagh and Saunders [10], suppose the problem variables are divided into three sets of variables: 1. The K basic variables, which have values strictly between their upper and lower bounds 2. A set of S superbasic variables, which have values strictly between their upper and lower bounds but are not basic 3. The remaining set of N ⫺ K ⫺ S nonbasic variables whose values lie on one or the other of the bounds The significance of this subdivision lies in the fact [a discovery attributed to A. Jain (see ref. 10)] that if a nonlinear program has NL nonlinear variables, then an optimal solution can be found at which the number of superbasic variables S is less than or equal to NL. This result suggests that it is sufficient to form the search direction d only from the superbasic variables. The concept of superbasic variable, therefore, effectively reduces the dimensionality of the search space d and, as a result, the size of any quasiNewton update formula that is stored and used to modify d. The nonbasic variables need in principle be considered only when no further improvement is possible within the space of the superbasics. Thus, if d ⫽ 0 (within some tolerance), then the reduced gradient with respect to the nonbasic variables is computed to determine if further improvement is possible by allowing one (or more) of these variables to leave their bounds and thus to join the set of superbasics. This strategy has led to the very efficient reduced gradient implementation MINOS [20] for linearly constrained problems and is an important concept in GRG implementations for large-scale problems [21, 22]. Computation of Derivatives. The GRG algorithm, and in fact all of the linearization-based algorithms require values of the derivatives of the problem functions. As in the unconstrained case, these derivates may be supplied or generated in various ways:

1. 2. 3. 4.

Supplied in analytical form by the user Generated by a symbolic manipulation routine Obtained from automatic differentiation routines Computed numerically by difference

Since the manual generation of analytical expressions for derivatives of functions that arise in engineering applications is very error prone, the last three alternatives are to be preferred. Algebraic modeling systems which allow the user to input model functions as algebraic expressions generally have capabilities to generate derivatives of these functions by interpreting their symbolic

430


form and then producing the required partial derivatives by applying calculus differentiation rules. In the case of optimizers that require the problem functions to be provided in the form of FORTRAN or C code, only the last two alternatives remain. In the case of models in which the objective and constraint functions can be explicitly written as expressions, automatic differentiation (AD) routines can generate derivative expressions and augment the user-supplied function routines with partial derivative expressions for all functions. A public source of such automatic differentiation codes is Argonne National Laboratory [23]. However, if none of the above three choices are available or if the optimization model consists of a set of procedures in which the constraint and objective functions are not explicit, then numerical differencing is the court of last resort. In optimization applications two forms of numerical differencing are in use: forward- or central-difference formulas. Since forward differencing requires N additional evaluations of each function, while central differencing requires 2N, the former is normally selected because the higher accuracy offered by central differencing is not really necessary. To further save function evaluations, two devices have been reported. Gabriele and Ragsdell [24] suggest the use of a test to determine whether a variable has changed sufficiently before the partial derivatives of the constraint functions with respect to that variable are computed. Given a differencing parameter ␦, the function values h˜ (i) k ⬅ hk(x1, . . . , xi⫺1, xi ⫹ ␦, si⫹1, . . . , xN)

k ⫽ 1, . . . , K

needed to estimate the partial derivatives ⭸hk / ⭸xi by forward difference, h˜ (i) ⫺ hk(t) ⭸ht)k ⯝ k ⭸xi ␦

(9.20)

are evaluated only if

冏

冏

(t⫺1) x (t) i ⫺ xi x (t⫺1) i

is greater than some selected tolerance ␧. This, of course, assumes that the value of each partial derivative with respect to xi is dominated by the variable xi. Such an approximation can be rather poor, for instance in the case h(x) ⫽ x1x2. Objective function evaluations can be saved in calculating the reduced gradient by taking advantage of the fact that changes in the basic variables can be estimated from changes in the nonbasic variables by using the linearization

9.3


431

dˆ ⫽ ⫺J⫺1Cd The procedure proposed by Gabriele and Ragsdell [24] calculates the reduced gradient at a point x (t) as follows. Given the differencing increment ␦: 1. Calculate the matrix J⫺1C at x (t) and denote the columns of this array Di. 2. Evaluate the objective function at x (t), that is, ƒ(t) ⫽ ƒ(x (t)). 3. For i ⫽ 1, . . . , N ⫺ K, set zi ⫽ xi ⫹ ␦

(9.21)

w ⫽ xˆ ⫺ ␦Di

(9.22)

ƒ˜ (i) ⫽ ƒ(w, x1, . . . , xi⫺1, zi, xi⫹1, . . . , xN⫺K)

(9.23)

Calculate

Evaluate

Then estimate the reduced gradient with respect to xi using (ⵜƒ˜ (t))i ⫽

ƒ˜ (i) ⫺ ƒ(t) ␦

(9.24)

The sequence of calculations (9.21)–(9.24) is executed N ⫺ K times to obtain ⵜƒ˜ (t). Thus, the number of objective function evaluations is N ⫺ K ⫹ 1 rather than the N ⫹ 1 that would be required if the complete gradient of ƒ were estimated. When K is large or ƒ(x) is complex to evaluate, this can be a significant savings. A potential disadvantage is that the inverse J⫺1 and the product J⫺1C must be available explicitly. Generation of Feasible Starting Points. The GRG algorithm requires that a

feasible solution be provided in order to initiate the algorithm logic. While feasible initial solutions can at times be provided by the user on the basis of physical knowledge of the system under consideration or, in some applications, by virtue of the specific structure of the problem, in general this is not the case. A more typical situation is that the user can provide an estimate of an initial solution but at that solution one or more of the constraints are violated. As in the linear programming case, the GRG algorithm can itself be used to correct this initial but infeasible solution either by augmenting the true objective function with a penalty term that includes the violated con-

432


straints or by first solving the equivalent of a phase I problem. The latter is usually to be preferred with NLP problems so as to avoid creating a complex and nonconvex objective function. Thus, for the general NLP problem, nonnegative artificial variables are added, one to each of the violated inequality constraints and unrestricted artificial variables are added, one to each violated equality constraints. An objective function is formed which consists of the sum of the artificial variables of the inequality constraints and the sum of the squares of the artificial variables of the violated equality constraints. The GRG algorithm is then applied to minimize this phase I objective subject to the original constraints, augmented by the artificial variables as appropriate. If at convergence of phase I optimization the objective function value is zero, the resulting solution will be feasible and can be used to initiate phase II or optimization of the actual objective function. If the converged phase I solution results in a nonzero objective function, it is likely that the constraint set is inconsistent. Example 9.12

Consider the problem of Example 9.11 with the initial solution x 0 ⫽ (4, 4). While the bounds are satisfied, both inequality constraints are violated. The phase I problem is thus formulated as follows: Minimize s1 ⫹ s2 Subject to g1(x, s) ⫽ 2x1 ⫺ x22 ⫹ s1 ⫺ 1 ⭓ 0 g2(x, s) ⫽ 9 ⫺ 0.8x21 ⫺ 2x2 ⫹ s2 ⭓ 0 4 ⭓ (x1, x2) ⭓ 0 (s1, s2) ⭓ 0 where (x, s)0 ⫽ (4, 4, 9, 11.8) is a feasible solution for the phase I problem. The GRG algorithm can readily be applied to solve this problem to yield a solution with (s1, s2) ⫽ 0. In contrast, if the constant in constraint g2 is reduced from 9 to 0.1, the phase I optimization will terminate with (s1, s2) ⫽ (0, 0.1), indicating an inconsistent constraint set. As can be seen from Figure 9.8, with the constant at 0.1, constraint g2 slides left, beyond the intersection of g1 with the x1 axis.

9.4

DESIGN APPLICATION

In this section we consider the least-cost design of multiproduct processes consisting of a series of batch and semicontinuous units. We first define the problem, then discuss its formulation, and finally consider its solution using a GRG code.

9.4

9.4.1

DESIGN APPLICATION

433

Problem Statement

The series process shown in Figure 9.10, consisting of three batch units (two reactors and a dryer) and five semicontinuous units (three pumps, one heat exchanger, and one centrifuge), is to process three products over a planning horizon of one year (8000 hr). The production requirements are as follows: Production Requirements, lb/yr Product 1 Product 2 Product 3

400,000 300,000 100,000

The batch processing times of each product in each of the three batch units are as follows: BATCH PROCESSING TIME, Product 1 Product 2 Product 3

HR

Reactor 1

Reactor 2

Dryer

3 6 2

1 — 2

4 8 4

All three products are to be processed serially in all units, with the exception that product 2 will by-pass reactor 2 and be sent directly to the centrifuge. The volume of material that must be processed at each stage to produce a pound of final product varies from product to product. The volume factors for the three products and eight units are given in Table 9.1.

Figure 9.10. Typical batch semicontinuous process.

434


Table 9.1

Volume Factors, ft3 / lb Final Product

Product 1

Product 2

Product 3

Batch Unit Volume Factors Sip Reactor 1 Reactor 2 Dryer

1.2 1.4 1.0

1.5 — 1.0

1.1 1.2 1.0

Continuous Unit Volume Factors Skp Pump 1 Pump 2 Heat exchanger Pump 3 Centrifuge

1.2 1.2 1.2 1.4 1.4

1.5 1.5 1.5 — 1.5

1.1 1.1 1.1 1.2 1.2

The plant equipment is to be sized for minimum capital cost. For batch units the equipment cost is given by a power law correlation of the form aiV␣i i, where Vi is the unit volume (ft3) and ai, ␣i are empirical constants. Semicontinuous unit costs are given by bkRk␤k, where Rk is the production rate (ft3 /hr) of unit k. The correlation coefficients for the eight equipment types are given in Table 9.2. 9.4.2

General Formulation

Many fine chemicals, pharmaceuticals, polymers, and foods are produced in processes that consist of a series of batch and/or semicontinuous processing stages. A batch operations is simply one in which a specified quantity of material is loaded into the unit, the material is subjected to some chemical or physical treatment for a specified time period, and then the resulting product is discharged. A semicontinuous unit is one that operates with continuous feed and removal of material but is intermittently started up or shut down. Table 9.2

Cost Coefficients for Eight Equipment Types

Equipment Type Reactor 1 Reactor 2 Dryer Pump 1 Pump 2 Heat exchanger Pump 3 Centrifuge

Batch Unit

Continuous Unit

ai

␣i

592 582 1200

0.65 0.39 0.52

bi

␤i

370 250 210 250 200

0.22 0.40 0.62 0.4 0.83

9.4

DESIGN APPLICATION

435

Batch/semicontinuous plants are normally used to produce multiple products so that the products actually share in the utilization of the plant equipment. It is preferable to operate such plants in the overlapping mode. Under such an operating policy each product is produced in campaigns that each involve the production of a certain number of batches of that product. During a campaign the process is dedicated to the exclusive production of one product. The processing of successive batches of that product is timed so that a new batch is started up as soon as a unit becomes available. Given a system consisting of three sequential batch stages with processing times t1, t2, t3, the operation of the process can be depicted in a bar graph (or Gantt chart), as shown in Figure 9.11. In this chart a solid horizontal line indicates that that unit is busy during the corresponding time interval. Note that unit 2 is occupied continuously. Unit 1 is idle two time units and unit 3 four time units between successive batches because they must wait before batches of material can be transferred to or from unit 2. For fixed processing times, each product will have associated with it a characteristic time, called the cycle time, between the completion of batches of that material. The cycle time will be equal to the maximum of the individual unit processing times, or Tp ⫽ max(ti) i

For the operation depicted in Figure 9.11, the cycle time will be equal to

Figure 9.11. Gantt chart for three-stage process.

436


Tp ⫽ max(t1, t2, t3) ⫽ t2 since t2 is the longest processing time. The product cycle time determines the batch size Bp and the number of batches Np that must be processed to produce a given quantity of material Qp over the fixed total production time T. Thus, Np ⫽ T/Tp and Bp ⫽ Qp /Np. If P products share in the use of the plant, then the total time T must be divided among these products. Thus,

冘 QB T ⭐ T p

p

(9.25)

p

p⫽1

p

The processing time of the semicontinuous units also influences the cycle time. First, if a batch unit i is preceded by semincontinuous unit l and followed by unit k, and if ␪kp indicates the processing time of unit k, then the total processing time of batch unit i, ␪ip, for a given product p becomes ␪ip ⫽ ␪lp ⫹ tip ⫹ ␪kp

(9.26)

Thus, the cycle time for product p will be given by Tp ⫽ max{max(␪ip), max(␪kp)} i

(9.27)

k

The volume of material that must be processed at stage i to produce a batch of Bp pounds of product p is equal to SipBp. Since this volume will be different for different products, the unit size must be selected to be big enough for all products. Thus, Vi ⫽ max(SipBp)

(9.28)

p

The continuous unit processing times ␪kp will be related to the batch size and the unit rate Rk (ft3 /hr) through the equation ␪kp ⫽

SkpBp Rk

(9.29)

Finally, the operating times of consecutive semicontinuous units must be related to each other. If for product p two semicontinuous units k and l are consecutive with l following k, then ␪kp ⭓ ␪lp

(9.30)

The cost function for the plant will consist of the sum of the capital costs of the batch and semicontinuous units. Thus,

9.4

Capital cost ⫽

DESIGN APPLICATION

冘a V ⫹ 冘b R i

␣i i

i

k

␤k k

437

(9.31)

k

The design optimization problem thus consists of minimizing (9.31) subject to Eqs. (9.25)–(9.30) plus nonnegativity conditions on the variables. This is an NLP problem with equality and inequality constraints. It could certainly be solved using a GRG algorithm. Note, however, that constraints (9.27) and (9.28) are nondifferentiable and consequently must be modified before a GRG algorithm can be employed. This is easily done by replacing (9.27) with the inequalities Tp ⭓ ␪ip

(9.32)

Tp ⭓ ␪kp

(9.33)

and

Similarly, (9.28) is replaced with the inequalities Vi ⭓ SipBp

(9.34)

As further simplification we can use (9.26) and (9.29) to eliminate ␪ip and ␪kp from the problem formulation, to result in a problem in the variables Vi, Rk, Tp, and Bp. 9.4.3

Model Reduction and Solution

For the process of Figure 9.10, the design optimization problem thus takes the form Minimize

a1V␣1 1 ⫹ a2V␣2 2 ⫹ a3V␣3 3 ⫹ b1R␤1 1 ⫹ b2R␤2 2 ⫹ b3R␤3 3 ⫹ b4R␤4 4 ⫹ b5R␤5 5

where the coefficients ai, ␣i, bi and ␤i are given Subject to

Total time constraints (9.25)

400,000 300,000 100,000 T1 ⫹ T2 ⫹ T3 ⭐ 8000 B1 B2 B3 Reactor 1 volume limits (9.34): V1 ⭓ 1.2B1

V1 ⭓ 1.5B2

V1 ⭓ 1.1B3

(9.35)

438


Reactor 2 volume limits (9.34): V2 ⭓ 1.4B1

V2 ⭓ 1.2B3

(9.36)

Dryer volume limits (9.34): V3 ⭓ 1.0B1

V3 ⭓ 1.0B2

V3 ⭓ 1.0B3

Cycle time of product 1, (9.25): T1 ⭓

1.2B1 R1

T1 ⭓

T1 ⭓

1.4B1 R4

1.2B1 R2

T1 ⭓

T1 ⭓

1.4B1 R5

1.2B1 R3 (9.37)

Cycle time of product 2 (9.33): T2 ⭓

1.5B2 R1

T2 ⭓

1.5B2 R2

T2 ⭓

1.5B2 R3

T2 ⭓

1.5B2 R5

Cycle time of product 3, (9.33): T3 ⭓

1.1B3 R1

T3 ⭓

T3 ⭓

1.2B3 R4

1.1B3 R2 T3 ⭓

T3 ⭓

1.1B3 R3

1.2B3 R5

Consecutive semicontinuous units 2 and 3 [(9.30) and (9.29)]: Product 1: Product 2: Product 3:

冧

1.2B1 1.2B1 ⭓ R2 R3 1.5B2 1.5B2 reduces to R ⭓ R ⭓ 3 2 R2 R3 1.1B3 1.1B3 ⭓ R2 R3

Consecutive semicontinuous units 4 and 5 [(9.30) and (9.29)]:

(9.38)

9.4

Product 1: Product 3:

DESIGN APPLICATION

冧

1.4B1 1.4B1 ⭓ R4 R5 reduces to R5 ⭓ R4 1.2B3 1.2B3 ⭓ R4 R5

439

(9.39)

For product 2, the heat exchanger will feed directly to the centrifuge, 1.5B2 1.5B2 ⭓ R2 R5

or

R5 ⭓ R2

(9.40)

Cycle time of product 1 [(9.32), (9.26), and (9.29)]: Reactor 1: Reactor 2: Dryer:

1.2B1 1.2B1 ⫹3⫹ R1 R2 1.2B1 1.4B1 T1 ⭓ ⫹1⫹ R3 R4 1.4B1 T1 ⭓ ⫹4 R5 T1 ⭓

(9.41)

Cycle time of product 2 [(9.32), (9.26), and (9.29)]: Reactor 1: Dryer:

1.5B2 1.5B2 ⫹6⫹ R1 R2 1.5B2 T2 ⭓ ⫹8 R5 T2 ⭓

Cycle time of product 3 [(9.32), (9.26), (9.29)]: Reactor 1: Reactor 2: Dryer:

1.1B3 1.1B3 ⫹2⫹ R1 R2 1.1B3 1.2B3 T3 ⭓ ⫹2⫹ R3 R4 1.2B3 T3 ⭓ ⫹4 R5 T3 ⭓

The problem thus has the following variables: three each of Tp, Bp, and Vi and five rates Rk. Since pump 2 and the heat exchanger are always operated together, there is no loss in setting R3 ⫽ R2 and eliminating (9.38) and R3 from the problem. The problem is all inequality constrained, with the following breakdown:

440

1 8 9 14 1 1


constraint of type (9.35) constraints of type (9.36) constraints of type (9.41) constraints of type (9.37) constraint of type (9.39) constraint of type (9.40)

Note, however, that the constraints of type (9.37) are really subsumed by the constraints of type (9.41). Thus, the problem can be reduced to 13 variables and 19 inequality constraints. A problem of this moderate size can readily be solved by the GRG2 code accessible through the standard version of Microsoft Excel Solver. Implementation only requires definition of the 19 constraints and objective functions via the Excel formula bar. For a discussion of the use of the Excel Solver, the reader is invited to consult [25]. Starting from an initial feasible solution with objective function value of 25, the optimal objective function value of $159,109.8 is obtained in just 30 iterations with convergence and precision parameter values of 10⫺6. The initial solution estimate and the optimal solution are given in Table 9.3. At the optimum, the cycle time of product 1 is limited by reactor 1, the cycle time of product 2 is limited by the dryer, and the cycle time of product 3 is limited by reactor 2. The semicontinuous times constitute a substantial fraction of the cycle time: 41 percent for product 1, 25 percent for product 2, and 44 percent for product 3. For a more detailed discussion of this class of design problems, the reader is directed to references 26 and 27.

Table 9.3

Results

Initial Estimate 3

Volume reactor 1, ft Volume reactor 2, ft3 Volume dryer, ft3 Rate of pump 1, ft3 / hr Rate of pump 2, ft3 / hr Rate of pump 4, ft3 / hr Rate of centrifuge, ft3 / hr Batch size, lb Product 1 Product 2 Product 3 Cycle time, hr Product 1 Product 2 Product 3

Optimal Solution

2000 2000 2000 1000 1000 1000 1000

1181.6 1225.2 875.2 1312.7 350.1 437.6 437.6

1000 1000 1000

875.2 787.7 875.2

10.0 10.0 10.0

6.800 10.700 7.150

REFERENCES

9.5

441

SUMMARY

In this chapter we have focused on methods that use linearizations of the nonlinear problem functions to generate good search directions. We examined two types of algorithms: feasible direction methods and extensions of the linear simplex method, which eventually led to the GRG algorithm. The family of feasible direction methods required the solution of an LP subproblem to determine a direction that was both a descent direction and strictly feasible. The formulation of the direction-finding subproblem uses an active constraint set strategy that could be susceptible to jamming unless special precautions such as the ␧-perturbation device were incorporated. This family of methods is suitable for nonconvex problems, generates a sequence of feasible points, but cannot directly accommodate nonlinear equality constraints. We found that the major disadvantage of feasible direction methods is the continued necessity of solving a series of LP subproblems. The second family of methods considered in this chapter seeks to avoid the LP solution by merely solving a set of linear equations to determine a good descent direction. For linearly constrained problems, the directions obtained will also be feasible. When only one nonbasic variable is used to define the search direction, we obtain the convex simplex method, the direct analog to the linear simplex method. When all nonbasic variables are used to define the direction, the reduced gradient method is obtained. Both variants essentially possess the convergence properties of Cauchy’s method. However, both can be accelerated by incorporating conjugate direction or quasi-Newton constructions. The reduced gradient strategy was then extended to accommodate nonlinear constraints. In this case the directions generated were no longer feasible, and hence the major revision required was the addition of a Newton iteration to adjust the basic variable to achieve feasibility. The resulting GRG algorithm could also be modified to incorporate quasi-Newton acceleration constructions and to treat variable bounds implicitly. The key assumptions underlying the GRG algorithm were continuous differentiability of the problem functions and the existence of nondegenerate, nonsingular bases throughout the iteration process. Implementation of the algorithm was found to involve a number of numerical issues whose efficient resolution could materially affect algorithm performance. The implementation of a GRG code, a number of which are in wide use, was thus seen to be a significant effort.

REFERENCES 1. Zoutendijk, G., Methods of Feasible Directions, Elsevier, Amsterdam, 1960. 2. Zangwill, W. I., Nonlinear Programming: A Unified Approach, Prentice-Hall, Englewood Cliffs, NJ, 1969.

442


3. Wolfe, P., ‘‘On the Convergence of Gradient Methods Under Constraints,’’ IBM J. Res. Develop., 16, 407–411 (1972). 4. Topkis, D. M., and A. F. Veinott, ‘‘On the Convergence of Some Feasible Direction Algorithms for Nonlinear Programming,’’ SIAM J. Control, 5, 268–279 (1967). 5. Zangwill, W. I., ‘‘The Convex Simplex Method,’’ Manag. Sci. A, 14, 221–238 (1967). 6. Wolfe, P., ‘‘Methods of Nonlinear Programming,’’ in Recent Advances in Mathematical Programming (R. L. Graves and P. Wolfe, Eds.), McGraw-Hill, New York, pp. 76–77, 1963. 7. McCormick, G. P., ‘‘Anti-Zig-Zagging by Bending,’’ Manag. Sci., 15, 315–320 (1969). 8. Bazaraa, M., and C. M. Shetty, Nonlinear Programming: Theory and Algorithms, Wiley, New York, 1979. 9. Hans-Tijan, T. Y., and W. I. Zangwill, ‘‘Analysis and Comparison of the Reduced Gradient and the Convex Simplex Method for Convex Programming,’’ paper presented at ORSA 41st National Meeting, New Orleans, Apr. 1972. 10. Murtagh, B. A., and M. A. Saunders, ‘‘Large Scale Linearly Constrained Optimization,’’ Math. Prog., 14, 41–72 (1978). 11. Abadie, J., and J. Carpentier, ‘‘Generalization of the Wolfe Reduced Gradient Method to the Case of Nonlinear Constraints,’’ in Optimization (R. Fletcher, Ed.), Academic, New York, 1969, Chapter 4. 12. Johnson, R. C., ‘‘A Method of Optimum Design,’’ ASME J. Mech. Des., 101, 667– 673 (1979). 13. Beightler, C., and D. J. Wilde, Foundation of Optimization Theory, Prentice-Hall, Englewood Cliffs, NJ, 1967. 14. Wilde, D. J., ‘‘Differential Calculus in Nonlinear Programming,’’ Oper. Res., 10, 764–773 (1962). See also Oper. Res. 13, 848–855 (1965). 15. Reklaitis, G. V., and D. J. Wilde, ‘‘Necessary Conditions for a Local Optimum without Prior Constraint Qualifications,’’ in Optimizing Methods in Statistics (J. S. Rustagi, Ed.), Academic, New York, 1971, pp. 339–368. 16. Lasdon, L. S., A. D. Waren, A. Jain, and M. Ratner, ‘‘Design and Testing of a Generalized Reduced Gradient Code for Nonlinear Programming,’’ ACM Trans. Math. Software, 4, 34–50 (1978). 17. Smith, S., and L. Lasdon, ‘‘Solving Large Sparse Nonlinear Programs Using GRG,’’ ORSA J. Comput., 4, 3–15 (Winter 1992). 18. Smeers, Y., ‘‘Generalized Reduced Gradient Method as an Extension of Feasible Direction Methods,’’ J. Opt. Theory Appl., 22, 209–226 (1977). 19. Luenberger, D. G., Introduction to Linear and Nonlinear Programming, 2nd ed., Addison-Wesley, Reading, MA, 1984. 20. Murtagh, B. A., and M. A. Saunders, ‘‘MINOS User’s Guide,’’ Tech. Rep. P3-20, Systems Optimization Laboratory, Stanford University, Stanford, CA, Feb. 1987. 21. Lasdon, L. S., and A. D. Warren, ‘‘Generalized Reduced Gradient Software for Linearly and Nonlinearly Constrained Problems,’’ in Design and Implementation of Optimization Software (H. Greenberg, Ed.), Sijthoff and Noordhoff, Holland, 1978. 22. Drud, A., ‘‘CONOPT—a Large Scale GRG Code,’’ ORSA J. Comput., 6 (Spring 1994).

PROBLEMS

443

23. Mathematics and Computer Science Division, Argonne National Laboratories, Argonne, IL, www.mcs.gov. 24. Gabriele, G. A., and K. M. Ragsdell, ‘‘The Generalized Reduced Gradient Method: A Reliable Tool for Optimal Design,’’ ASME J. Eng. Ind., 99, 384–400 (1977). 25. Fylstra, D., L. Lasdon, and A. Waven, ‘‘Design and Use of the Microsoft Excel Solver’’, Interfaces, 28, 29–55 (Sept.–Oct. 1998). 26. Knopf, F. C., M. R. Okos, and G. V. Reklaitis, ‘‘Optimal Design of Batch / SemiContinuous Processes,’’ Ind. Eng. Chem. Proc. Des. Dev., 21, 79–86 (1982). 27. Biegler, L. T., and I. E. Grossmann, and A. W. Westerberg, Systematic Methods of Chemical Process Design, Prentice-Hall, Upper Saddle River, NJ, 1997.

PROBLEMS

9.1. In the development of the feasible direction method, descent directions are employed that remain in the interior of the feasible region. Is this a good idea? Why or why not? 9.2. Can the feasible direction method be applied to equality-constrained problems? Explain. 9.3. (a) What is the major advantage of using an active constraint strategy? (b) How would such a strategy be used with the GRG method? 9.4. What are the major simplifications associated with a linear constraint set when using the GRG method? 9.5. The feasible direction method is defined with equal weighting of feasibility and descent. Formulate a variant that might weigh descent more heavily. (Hint: Multiply ␪ by suitable constants.) 9.6. What potential difficulties can arise if a step of the convex simplex method is executed with a degenerate basis? 9.7. Describe a procedure for estimating the remaining objective function improvement that could be carried out as part of the convex simplex algorithm for problems with convex objective functions. 9.8. Outline the way in which the conjugate directions method could be incorporated within the convex simplex method. 9.9. Under what conditions would a SLP solution approach be preferable to using a GRG solution approach? 9.10. Could the convex simplex method find a solution point which is interior to the linear constraint set? Explain. 9.11. For an inequality-constrained problem, what is the main computational difference between the GRG algorithm and the method of feasible directions?

444


9.12. In the GRG method when a quasi-Newton update is used, it is only applied to the components of the direction vector corresponding to the nonbasic variables? Why is it not applied to the complete direction vector? 9.13. Outline a modification of the convex simplex method in which variable bounds ai ⭐ xi ⭐ bi would be handled without treating them as explicit inequality constraints. 9.14. Consider the problem Minimize

ƒ(x) ⫽ 100(x2 ⫺ x21)2 ⫹ (1 ⫺ x21)

Subject to g1(x) ⫽ x1 ⫹ 1 ⭓ 0 g2(x) ⫽ 1 ⫺ x2 ⭓ 0 g3(x) ⫽ 4x2 ⫺ x1 ⫺ 1 ⭓ 0 g4(x) ⫽ 1 ⫺ 0.5x1 ⫺ x2 ⭓ 0 Given the point x 0 ⫽ (⫺1, 1). (a) Perform one full iteration of the convex simplex method. (b) Perform one full iteration of the reduced gradient method. (c) Sketch the feasible region and explain the difference in the results. 9.15. Given the linearly constrained NLP problem Minimize

ƒ(x) ⫽ (x1 ⫺ 2)2 ⫹ (x2 ⫺ 2)2

Subject to g1(x) ⫽ 3 ⫺ x1 ⫺ x2 ⭓ 0 g2(x) ⫽ 10x1 ⫺ x2 ⫺ 2 ⭓ 0 x2 ⭓ 0 and the starting point x 0 ⫽ (0.2, 0). (a) Carry out three iterations of the method of feasible directions. (b) Carry out three iterations of the reduced gradient method. In both cases, carry out subproblem solutions using graphical or algebraic shortcuts. (c) Constrast the iteration paths of the two methods using a graph. 9.16. Suppose the constraint g1 of problem 9.15 is replaced by g1(x) ⫽ 4.5 ⫺ x12 ⫺ x22 ⭓ 0

PROBLEMS

445

(a) Discuss how this modification will affect the iteration path of problem 9.15(b). (b) Calculate the reduced gradient at the feasible intersection of constraints g1 and g2. Is the point a local minimum? (c) Calculate the search direction from the point of part (b) using the method of feasible directions. Contrast to the direction generated by the GRG method. (d) Calculate the reduced gradient at the point (–32 , –32 ). Is this point a local minimum? 9.17. Suppose the constraint g1 of problem 9.15 is replaced by g1(x) ⫽ 2.25 ⫺ x1x2 ⭓ 0 Reply to questions (b), (c), and (d) of problem 9.16 with this problem modification. 9.18. Consider the problem Minimize

ƒ(x) ⫽ (x1 ⫺ x2)2 ⫹ x2

Subject to g1(x) ⫽ x1 ⫹ x2 ⫺ 1 ⭓ 0 g2(x) ⫽ 3 ⫺ x1 ⫹ 2x2 ⭓ 0 x1 ⭓ 0

x2 ⭓ 0

with initial point x 0 ⫽ (3, 0). (a) Calculate the search direction at x 0 using the convex simplex method, the reduced gradient method, and the method of feasible directions. (b) Use a sketch to show the differences in the search directions. 9.19. Consider the problem Minimize

ƒ(x) ⫽ x1 ⫹ x2

Subject to h(x) ⫽ x21 ⫹ x22 ⫺ 1 ⫽ 0 (a) Calculate the reduced gradient at the point x 0 ⫽ (0, 1). (b) Suppose the algorithm progresses from x 0 to the point (⫺1, 0). Which variable should be selected to be independent and which dependent? Why?

446


9.20. Consider the problem Maximize

ƒ(x) ⫽ 2x1 ⫺ 3(x2 ⫺ 4)2

Subject to g1(x) ⫽ 10 ⫺ x12 ⫺ x22 ⭓ 0 g2(x) ⫽ 9 ⫺ x21 ⫺ (x2 ⫺ 4)2 ⭓ 0 x1 ⭓ 0 (a) Given the starting point x 0 ⫽ (2, 2): (1) Calculate the next iteration point that would result if a method of feasible direction were applied with ␧ ⫽ 10⫺3. (2) Calculate the next iteration point that would be attained if the generalized reduced gradient method were applied with ␧ ⫽ 10⫺3. (3) Exhibit these results graphically. (b) Given the point (0, 1), repeat the three parts outlined in (a). (c) Use the reduced gradient to determine whether the point (1, 3) is a local maximum. How can you prove if it is the global maximum? 9.21. Consider the problem Minimize

ƒ(x) ⫽ ⫺x1x22x33 /81

Subject to h1(x) ⫽ x13 ⫹ x22 ⫹ x3 ⫺13 ⫽ 0 h2(x) ⫽ x22x3⫺1 / 2 ⫺ 1 ⫽ 0 (a) Calculate the reduced gradient at the point x ⫽ (1, 兹3, 9). (b) Show that this point is a constrained local minimum. (c) Confirm (b) by using the Lagrangian conditions. 9.22. Consider the problem Minimize

ƒ(x) ⫽ ⫺x1 ⫺ 2x2 ⫺ 4x1x2

Subject to g1(x) ⫽ 9 ⫺ (x1 ⫺ 2)2 ⫺ (x2 ⫺ 1)2 ⭓ 0 g2(x) ⫽ x1 ⭓ 0 g3(x) ⫽ x2 ⭓ 0 (a) Caculate the direction that would be selected by the feasible directions method at x 0 ⫽ (0, 0). (b) Repeat (a) with x 0 ⫽ (5, 1).

PROBLEMS

447

(c) If the direction calculated in (a) were used, what would be the next iteration point? 9.23. Given the NLP problem Minimize

ƒ(x) ⫽ (x1 ⫺ 1)2 ⫹ (x2 ⫺ 1)2 ⫹ (x3 ⫺ 1)2

Subject to g1(x) ⫽ x1 ⫹ x2 ⫹ x3 ⫺ 2 ⭐ 0 h1(x) ⫽ 1.5x21 ⫹ x22 ⫹ x32 ⫺ 2 ⫽ 0 Consider a feasible point x ⫽ (1, –12 , –12 ). (a) Calculate the reduced gradient at the point x and show that this corresponds to a Kuhn–Tucker point. (b) Calculate the multipliers at x. (c) Suppose the constraint h1(x) is deleted. Calculate the reduced gradient and the new search direction. What would be the new point generated by the GRG algorithm? 9.24. (a) Your reduced gradient code used to solve the problem Minimize

ƒ(x) ⫽ 4x1 ⫺ x22 ⫹ x23 ⫺ 12

Subject to h1(x) ⫽ 20 ⫺ x12 ⫺ x22 ⫽ 0 h2(x) ⫽ x1 ⫹ x3 ⫺ 7 ⫽ 0 x1, x2, x3, ⭓ 0 terminates at the point (2.5, 3.71, 4.5). Is this the solution to the problem? Why? (b) Upon checking the problem formulation, you discover that the constraint h3(x) ⫽ x3x⫺1 1 ⫺ 2 ⫽ 0 was left out. What is the solution now? 9.25. Consider the problem Minimize

ƒ(x) ⫽ exp(x2) ⫹ x3 ⫺ exp(x4)

Subject to g(x) ⫽ exp(x1 ⫺ 1) ⫺ x2 ⫹ x3 ⫺ 1 ⭓ 0 g2(x) ⫽ x2 ⫺ x23 ⫹ x4 ⭓ 0 all xi ⭓ 0 Suppose the GRG method commences with the point (1, 0, 0, 0).

448


(a) Show that all choices of basic variables will be either degenerate or degenerate and singular. (b) Show that the choice x ⫽ (x1, x2) will lead to a search direction that will force violation of variable nonnegativity. (c) Suggest a way to calculate an acceptable descent direction. 9.26. Obtain a starting feasible solution to the optimization problem posed in problem 7.18 using the GRG code assessable via the Excel Solver. 9.27. Consider the problem Minimize

ƒ(x) ⫽ exp(⫺x1) ⫹ exp(x2 ⫺ 1) ⫺ exp(x3) ⫹ x4

Subject to h1(x) ⫽ exp(x1 ⫺ 1) ⫺ x2 ⫹ x3 ⫺ 1 ⫽ 0 h2(x) ⫽ x2 ⫺ (x3 ⫺ 1)2 ⫹ x4 ⫺ 2 ⫽ 0 x1, x3, x4 ⭓ 0

x2 ⭓ 1

point x* ⫽ (1, 1, 1, 1) (a) Explain why (x1, x3) is not a suitable choice of basic variables in calculating the reduced gradient at x*. (b) Explain why (x2, x3) is not a suitable choice of basic variables in calculating the reduced gradient at x*. (c) Choose an appropriate set of basic variables, calculate the reduced gradient at the point x*, and use it to determine a direction vector for search. Is x* the local minimum point? 9.28. Consider the problem Minimize

ƒ(x) ⫽ exp[(x1 ⫺ 1)2 ⫹ (x4 ⫺ 2)2] ⫺ 5 exp(⫺x3)

Subject to h1(x) ⫽ 4x12 ⫹ x2 ⫹ x32 ⫺ 4x42 ⫺ 4 ⫽ 0 xi ⭓ 0

i ⫽ 1, 4

point x 0 ⫽ (1, 4, 0, 1) (a) Compute the reduced gradient at x 0 and show that the point is not a local minimum. (b) Using the results of (a) give a direction vector d which will be a descent direction. Write out the equation of feasible points corresponding to this direction.

PROBLEMS

449

(c) Suppose the additional constraint g1(x) ⫽ 4(1 ⫺ x4) ⫺ (x1 ⫺ 1)2 ⭓ 0 is imposed on the problem. How does this affect the direction vector computed in (b)? What can you say about point x 0 under these conditions?

10 QUADRATIC APPROXIMATION METHODS FOR CONSTRAINED PROBLEMS In the preceding two chapters we considered a number of alternative strategies for exploiting linear approximations to nonlinear problem functions. In general we found that, depending upon the strategy employed, linearizations would either lead to vertex points of the linearized constraint sets or generate descent directions for search. In either case, some type of line search was required in order to approach the solution of non-corner-point constrained problems. Based upon our experience with unconstrained problems, it is reasonable to consider the use of higher order approximating functions since these could lead directly to better approximations of non-corner-point solutions. For instance, in the single-variable case we found that a quadratic approximating function could be used to predict the location of optima lying in the interior of the search interval. In the multivariable case, the use of a quadratic approximation (e.g., in Newton’s method) would yield good estimates of unconstrained minimum points. Furthermore, the family of quasiNewton methods allowed us to reap some of the benefits of a quadratic approximation without explicitly developing a full second-order approximating function at each iteration. In fact, in the previous chapter we did to some extent exploit the acceleration capabilities of quasi-Newton methods by introducing their use within the direction generation mechanics of the reduced gradient and gradient projection methods. Thus, much of the discussion of the previous chapters does point to the considerable algorithmic potential of using higher order, specifically quadratic, approximating functions for solving constrained problems. In this chapter we examine in some detail various strategies for using quadratic approximations. We begin by briefly considering the consequence of direct quadratic approximation, the analog of the successive LP strategy. Then we investigate the use of the second derivatives and Lagrangian con450


10.1

DIRECT QUADRATIC APPROXIMATION

451

structions to formulate quadratic programming (QP) subproblems, the analog to Newton’s method. Finally, we discuss the application of quasi-Newton formulas to generate updates of quadratic terms. We will find that the general NLP problem can be solved very efficiently via a series of subproblems consisting of a quadratic objective function and linear constraints, provided a suitable line search is carried out from the solution of each such subproblem. The resulting class of exterior point algorithms can be viewed as a natural extension of quasi-Newton methods to constrained problems.

10.1


Analogous to the successive LP strategy of Section 8.1, we could approach the solution of the general NLP problem by simply replacing each nonlinear function by its local quadratic approximation at the solution estimate x 0 and solving the resulting series of approximating subproblems. If each function ƒ(x) is replaced by its quadratic approximation q(x; x 0) ⫽ ƒ(x 0) ⫹ ⵜƒ(x 0)T (x ⫺ x 0) ⫹ –12 (x ⫺ x 0)T ⵜ2ƒ(x 0)(x ⫺ x 0) then the subproblem becomes one of minimizing a quadratic function subject to quadratic equality and inequality constraints. While it seems that this subproblem structure ought to be amendable to efficient solution, in fact, it is not. To be sure, the previously discussed strategies for constrained problems can solve this subproblem but at no real gain over direct solution of the original problem. For a sequential strategy using subproblem solutions to be effective, the subproblem solutions must be substantially easier to obtain than the solution of the original problem. Recall from the discussion in Section 9.2.3 that problems with a quadratic positive-definite objective function and linear constraints can be solved in a finite number of reduced gradient iterations provided that quasi-Newton or conjugate gradient enhancement of the reduced gradient direction vector is used. Of course, while the number of iterations is finite, each iteration requires a line search, which strictly speaking is itself not a finite procedure. However, as we will discover in Chapter 11, there are specialized methods for these so-called QP problems that will obtain a solution in a finite number of iterations without line searching, using instead simplex like pivot operations. Given that QP problems can be solved efficiently with truly finite procedures, it appears to be desirable to formulate our approximating subproblems as quadratic programs. Thus, assuming that the objective function is twice continuously differentiable, a plausible solution strategy would consist of the following steps:

452

QUADRATIC APPROXIMATION METHODS FOR CONSTRAINED PROBLEMS

Direct Successive Quadratic Programming Solution

Given x 0, an initial solution estimate, and a suitable method for solving QP subproblems. Step 1. Formulate the QP problem Minimize

ⵜƒ(x(t))Td ⫹ –12 d T ⵜ2ƒ(x(t))d

Subject to hk (x(t)) ⫹ ⵜhk (x(t))Td ⫽ 0 gj(x ) ⫹ ⵜgj(x ) d ⭓ 0 (t)

(t) T

k ⫽ 1, . . . , K j ⫽ j ⫽ 1, . . . , J

Step 2. Solve the QP problem and set x(t⫹1) ⫽ x(t) ⫹ d. Step 3. Check for convergence. If not converged, repeat step 1. As shown in the following example, this approach can be quite effective. Example 10.1

Solve the problem Minimize

⫺2 ƒ(x) ⫽ 6x1x⫺1 2 ⫹ x2x1

Subject to h(x) ⫽ x1x2 ⫺ 2 ⫽ 0 g(x) ⫽ x1 ⫹ x2 ⫺ 1 ⭓ 0 from the initial feasible estimate x 0 ⫽ (2, 1) using the direct successive QP strategy. At x 0, ƒ(x 0) ⫽ 12.25, h(x 0) ⫽ 0, and g(x 0) ⫽ 2 ⬎ 0. The derivatives required to construct the QP subproblem are ⫺3 ⫺2 ⫺2 T ⵜƒ(x) ⫽ ((6x⫺1 2 ⫺ 2x2x1 ), (⫺6x1x2 ⫹ x1 ))

ⵜ2ƒ ⫽

冉

(6x2x⫺4 1 ) ⫺3 (⫺6x⫺2 2 ⫺ 2x1 )

ⵜh(x) ⫽ (x2, x1)T

Thus, the first QP subproblem will be

冊

⫺3 (⫺6x⫺2 2 ⫺ 2x1 ) ⫺3 (12x1x2 )

10.1


冉

Minimize

23 47 1 T (–– 4 , ⫺ –– 4 )d ⫹ – 2d

Subject to

(1, 2)d

3 8 25 4

– ⫺––

冊

25 ⫺–– 4

24

453

d

⫽0

(1, 1)d ⫹ 2 ⭓ 0 Since the first constraint can be used to eliminate one of the variables, that is, d1 ⫽ ⫺ 2d2 the resulting single-variable problem can be solved easily analytically to give d 0 ⫽ (⫺0.92079, 0.4604) Thus, the new point becomes x(1) ⫽ x 0 ⫹ d 0 ⫽ (1.07921, 1.4604) at which point ƒ(x(1)) ⫽ 5.68779 h(x(1)) ⫽ ⫺0.42393 g(x(1)) ⬎ 0 Note that the objective function value has improved substantially but that the equality constraint is violated. Suppose we continue with the solution procedure. The next subproblem is

冉

冊

6.4595 ⫺4.4044 d 4.1579

Minimize

(1.78475, ⫺2.17750)d ⫹ –12 d T

Subject to

(1.4604, 1.07921)d ⫺ 0.42393 ⫽ 0

⫺4.4044

(1, 1)d ⫹ 1.5396 ⭓ 0 The solution is d (1) ⫽ (⫺0.03043, 0.43400), resulting in the new point, with x(2) ⫽ (1.04878, 1.89440) ƒ(x(2)) ⫽ 5.04401 h(x(2)) ⫽ ⫺0.013208 g(x(2)) ⬎ 0 Note that both the objective function value and the equality constraint violation are reduced. The next two iterations produce the results

454


x(3) ⫽ (1.00108, 1.99313)

ƒ(x(3)) ⫽ 5.00457

h(x(3)) ⫽ ⫺4.7 ⫻ 10⫺3

x(4) ⫽ (1.00014, 1.99971)

ƒ(x(4)) ⫽ 5.00003

h(x(4)) ⫽ ⫺6.2 ⫻ 10⫺6

The exact optimum is x* ⫽ (1, 2) with ƒ(x*) ⫽ 5.0; a very accurate solution has been obtained in four iterations. As is evident from Figure 10.1, the constraint linearizations help to define the search directions, while the quadratic objective function approximation effectively fixes the step length along that direction. Thus, the use of a nonlinear objective function approximation does lead to nonvertex solutions. However, as shown in the next example, the self-bounding feature of the quadratic approximation does not always hold.

Figure 10.1 Plot of Example 10.1

10.1


455

Example 10.2

Suppose the objective function and equality constraint of Example 10.1 are interchanged. Thus, consider the problem Minimize

ƒ(x) ⫽ x1x2

⫺2 Subject to h(x) ⫽ 6x1x⫺1 2 ⫹ x2x1 ⫺ 5 ⫽ 0

g(x) ⫽ x1 ⫹ x2 ⫺ 1 ⭓ 0 The optimal solution to this problem is identical to that of Example 10.1. With the starting point x 0 ⫽ (2, 1) the first subproblem becomes

冉冊

0 1 d 1 0

Minimize

(1, 2)d ⫹ –12 d T

Subject to

23 47 29 (–– 4 , ⫺–– 4 )d ⫹ –– 4 ⫽ 0

(1, 1)d ⫹ 2 ⭓ 0 The solution to this subproblem is d 0 ⫽ (⫺1.7571, ⫺0.24286), at which both constraints are tight. Thus, a subproblem corner point is reached. The resulting intermediate solution is x1 ⫽ x 0 ⫹ d 0 ⫽ (0.24286, 0.75714) with ƒ(x(1)) ⫽ 0.18388

h(x(1)) ⫽ 9.7619

g(x(1)) ⫽ 0

Although the objective function value decreases, the equality constraint violation is worse. The next subproblem becomes Minimize

冉冊

(0.75714, 0.24286)d ⫹ –12 d T

0 1 d 1 0

Subject to ⫺97.795d1 ⫹ 14.413d2 ⫹ 9.7619 ⫽ 0 d1 ⫹ d 2 ⭓ 0 The resulting new point is x(2) ⫽ (0.32986, 0.67015) with

456


ƒ(x(2)) ⫽ 0.2211

h(x(2)) ⫽ 4.1125

g(x(2)) ⫽ 0

Again, g(x) is tight. The next few points obtained in this fashion are x(3) ⫽ (0.45383, 0.54618) x(4) ⫽ (⫺0.28459, 1.28459) x(5) ⫽ (⫺0.19183, 1.19183) These and all subsequent iterates all lie on the constraint g(x) ⫽ 0. Since the objective function decreases toward the origin, all iterates will be given by the intersection of the local linearization with g(x) ⫽ 0. Since the slope of the linearization becomes larger or smaller than ⫺1 depending upon which side of the constraint ‘‘elbow’’ the linearization point lies, the successive iterates simply follow an oscillatory path up and down the surface g(x) ⫽ 0. Evidently the problem arises because the linearization cannot take into account the sharp curvature of the constraint h(x) ⫽ 0 in the vicinity of the optimum. Since both the constraint and the objective function shapes serve to define the location of the optimum, both really ought to be take into account. This can be accomplished implicitly by limiting the allowable step or by using a penalty function–based line search, as discussed in Section 8.1.2. Alternatively, the constraint curvature can be taken into account explicitly by incorporating it into a modified subproblem objective function, as discussed in the next section. 10.2 QUADRATIC APPROXIMATION OF THE LAGRANGIAN FUNCTION

The examples of the previous section suggest that it is desirable to incorporate into the subproblem definition not only the curvature of the objective function but also that of the constraints. However, based on computational considerations, we also noted that it is preferable to deal with linearly constrained rather than quadratically constrained subproblems. Fortunately, this can be accomplished by making use of the Lagrangian function, as will be shown below. For purposes of this discussion, we first consider only the equalityconstrained problem. The extension to inequality constraints will follow in a straightforward fashion. Consider the problem Minimize

ƒ(x)

Subject to h(x) ⫽ 0 Recall that the necessary conditions for a point x* to be a local minimum are that there exist a multiplier v* such that

10.2

QUADRATIC APPROXIMATION OF THE LAGRANGIAN FUNCTION

ⵜx L(x*, v*) ⫽ ⵜƒ* ⫺ v*T ⵜh* ⫽ 0

and

h(x*) ⫽ 0

457

(10.1)

Sufficient conditions for x* to be a local minimum (see Theorem 5.7 of Chapter 5) are that conditions (10.1) hold and that the Hessian of the Lagrangian function, ⵜ2x L(x*, v*) ⫽ ⵜ2ƒ* ⫺ (v)*T ⵜ2h*

satisfies d T ⵜ2x Ld ⬎ 0

for all d such that (ⵜh)*Td ⫽ 0

(10.2)

Given some point (x, v), we construct the following subproblem expressed in terms of the variables d: Minimize

ⵜƒ(x)Td ⫹ –21 d T ⵜx2 L(x, v)d

Subject to h(x) ⫹ ⵜh(x)Td ⫽ 0

(10.3) (10.4)

We now observe that if d* ⫽ 0 is the solution to the problem consisting of (10.3) and (10.4), then x must satisfy the necessary conditions for a local minimum of the original problem. First note that if d* ⫽ 0 solves the subproblem, then form (10.4) it follows that h(x) ⫽ 0; in other words, x is a feasible point. Next, there must exist some v* such that the subproblem functions satisfy the Lagrangian necessary conditions at d* ⫽ 0. Thus, since the gradient of the objective function (10.3) with respect to d at d* ⫽ 0 is ⵜƒ(x) and that of (10.4) is ⵜh(x), it follows that ⵜƒ(x) ⫺ (v)*T ⵜh(x) ⫽ 0

(10.5)

Clearly v* will serve as the Lagrange multiplier for the original problem, and thus x satisfies the necessary conditions for a local minimum. Now suppose we take the v* given above and reformulate the subproblem at (x, v*). Clearly, (10.5) still holds so that d ⫽ 0 satisfies the necessary conditions. Suppose the subproblem also satisfies the second-order sufficient conditions at d ⫽ 0 with v*. Then, it must be true that d{ⵜ2x L(x, v*) ⫺ v*(0)}d ⬎ 0

for all d such that ⵜh(x)Td ⫽ 0

Note that the second derivative with respect to d of (10.4) is zero, since it is a linear function in d. Consequently, the above inequality implies that d T ⵜ2x L(x, v*)d is positive also. Therefore, the pair (x , v*) satisfies the sufficient conditions for a local minimum of the original problem. This demonstration indicates that the subproblem consisting of (10.3) and (10.4) has the following very interesting features:

458


1. If no further corrections can be found, that is, d ⫽ 0, then the local minimum of the original problem will have been obtained. 2. The Lagrange multipliers of the subproblem can be used conveniently as estimates of the multipliers used to formulate the next subproblem. 3. For points sufficiently close to the solution of the original problem the quadratic objective function is likely to be positive definite, and thus the solution of the QP subproblem will be well behaved. By making use of the sufficient conditions stated for both equality and inequality constraints, it is easy to arrive at a QP subproblem formulation for the general case involving K equality and J inequality constraints. If we let L(x, u, v) ⫽ ƒ(x) ⫺

冘v h (x) ⫺ 冘u g (x) k

k

(10.6)

j j

then at some point (x, u, v) the subproblem becomes Minimize

q(d; x) ⬅ ⵜƒ(x)Td ⫹ –12 d T ⵜx L(x, u, v)d

Subject to h˜ k (d; x) ⬅ hk (x) ⫹ ⵜhk (x)Td ⫽ 0

(10.7)

k ⫽ 1, . . . , K (10.8a)

g˜ j(d; x) ⬅ gj(x) ⫹ ⵜgj(x)Td ⭓ 0

j ⫽ 1, . . . , J (10.8b)

The algorithm retains the basic steps outlined for the direct QP case. Namely, given an initial estimate x 0 as well as u0 and v0 (the latter could be set equal to zero, we formulate the subproblem [Eqs. (10.7), (10.8a), and (10.8b)]; solve it; set x(t⫹1) ⫽ x(t) ⫹ d; check for convergence; and repeat, using as next estimates of u and v the corresponding multipliers obtained at the solution of the subproblem. Example 10.3

Repeat the solution of the problem of Example 10.1 using the Lagrangain QP subproblem with initial estimates x 0 ⫽ (2, 1)T, u0 ⫽ 0, and v0 ⫽ 0. The first subproblem becomes

冉

Minimize

23 47 1 T (–– 4 , ⫺–– 4 )d ⫹ – 2d

Subject to

(1, 2)d

3 8 25 4

– ⫺ ––

冊

25 ⫺–– 4

24

d

⫽0

(1, 1)d ⫹ 2 ⭓ 0 This is exactly the same as the fist subproblem of Example 10.1, because with the initial zero estimates of the multipliers of the constraint terms of the Lagrangian will vanish. The subproblem solution is thus, as before,

10.2

459


d 0 ⫽ (⫺0.92079, 0.4604)T Since the inequality constraint is loose at this solution, u(1) must equal zero. The equality constraint multiplier can be found from the solution of the Lagrangian necessity conditions for the subproblem. Namely, ˜ 0; x 0) ⵜq(d 0; x 0) ⫽ v ⵜh(d

or

再

冉

23 47 T (–– 4 , ⫺–– 4 ) ⫹ d

⫺–83 25 ⫺–– 4

冊冎

25 ⫺–– 4

24

⫽ v(1, 2)T

Thus, v(1) ⫽ 2.52723. Finally, the new estimate of the problem solution will be x(1) ⫽ x 0 ⫹ d 0, or x(1) ⫽ (1.07921, 1.4604)T

ƒ(x(1)) ⫽ 5.68779

h(x(1)) ⫽ ⫺0.42393

as was the case before. The second subproblem requires the gradients ⵜƒ(x(1)) ⫽ (1.78475, ⫺2.17750)T

冉冉冊

ⵜh(x(1)) ⫽ (1.4604, 1.07921)T

冊

6.45954 ⫺4.40442 4.15790

ⵜ2ƒ(x(1)) ⫽

⫺4.40442

0 1 1 0

ⵜ2h(x(1)) ⫽

The quadratic term is therefore equal to ⵜ2L ⫽ ⵜ2ƒ ⫺ v ⵜ2h

冉冊冉

⫽ ⵜ2ƒ ⫺ 2.52723

冊

0 1 6.45924 ⫺6.93165 ⫽ 1 0 ⫺6.93165 4.15790

The complete subproblem becomes

冉

(1.78475, ⫺2.17750)d ⫹ –12 d T

Subject to

1.4604d1 ⫹ 1.07921d2 ⫽ 0.42393 d1 ⫹ d2 ⫹ 1.539604 ⭓ 0

冊

6.45924 ⫺6.93165 d 4.15790

Minimize

⫺6.93165

460


The solution is d (1) ⫽ (0.00614, 0.38450). Again, since g˜(d (1); x(1)) ⬎ 0, u ⫽ 0, and the remaining equality constraint multiplier can be obtained from (2)

ⵜq(d (1); x(1)) ⫽ vT ⵜh(d (1); x(1))

or

冉

冊冉

⫺0.84081 ⫺0.62135

⫽v

冊

1.46040 1.07921

Thus, v(2) ⫽ ⫺0.57574

x(2) ⫽ (1.08535, 1.84490)T

and

with ƒ(x(2)) ⫽ 5.09594

h(x(2)) ⫽ 2.36 ⫻ 10⫺3

and

Continuing the calculations for a few more iterations, the results obtained are x(3) ⫽ (0.99266, 2.00463)T

v(3) ⫽ ⫺0.44046

ƒ(3) ⫽ 4.99056

h(3) ⫽ ⫺1.008 ⫻ 10⫺2

and x(4) ⫽ (0.99990, 2.00017)T

v(4) ⫽ ⫺0.49997

ƒ(4) ⫽ 5.00002

h(4) ⫽ ⫺3.23 ⫻ 10⫺5

It is interesting to note that these results are essentially comparable to those obtained in Example 10.1 without the inclusion of the constraint secondderivative terms. This might well be expected, because at the optimal solution (1, 2) the constraint contribution to the second derivative of the Lagrangian is small: ⵜ2ƒ* ⫺ v* ⵜ2h* ⫽

冉

12

⫺3.5

冊冉冊冉冊冉

⫺3.5

1.5

⫺

⫺

1 2

冊

0 1 12 ⫺3.0 ⫽ 1 0 ⫺3.0 1.5

The basic algorithm illustrated in the preceding example can be viewed as an extension of Newton’s method to accommodate constraints. Specifically, if no constraints are present, the subproblem reduces to

10.2


461

Minimize ⵜƒ(x)Td ⫹ –12 d T ⵜ2ƒ(x)d which is the approximation used in Newton’s method. Viewing it in this light, we might expect that some of the difficulties associated with Newton’s method may also arise in the extension. Specifically, we might expect to generate nonminimum stationary points if the second derivative term is nonpositive definite and may expect failure to converge from poor initial solution estimates. As shown in the next example, this is in fact the case. Example 10.4

Consider the problem of Example 10.2 with the initial estimate x 0 ⫽ (2, 2.789) and u ⫽ v ⫽ 0. The first subproblem will be given by

冉冊

0 1 d 1 0

Minimize

(2.789, 2)d ⫹ –12 d T

Subject to

1.4540d1 ⫺ 1.2927d2 ⫽ 1.3 ⫻ 10⫺4 d1 ⫹ d2 ⫹3.789 ⭓ 0

The solution is d 0 ⫽ (⫺1.78316, ⫺2.00583). The inequality constraint is tight, so both constraint multipliers must be computed. The result of solving the system ˜ 0; x 0) ⫹ u ⵜg˜ (d 0; x 0) ⵜq˜ (d 0; x 0) ⫽ v ⵜh(d

冉

冊冉

冊冉冊

0.78317 ⫺145.98 1 ⫽v ⫹u 0.21683 19.148 1

is v(1) ⫽ ⫺0.00343 and u(1) ⫽ 0.28251. At the corresponding intermediate point, x(1) ⫽ (0.21683, 0.78317)T we have ƒ(x(1)) ⫽ 0.1698

h(x(1)) ⫽ 13.318

g(x(1)) ⫽ 0

Note that the objective function decreases substantially, but the equality constraint violation becomes very large. The next subproblem constructed at x(1) with multiplier estimates u(1) and v(1) is

462


Minimize

(0.78317, 0.21683)d

再冉冊冉

⫹ –12 d T

冊

冎

0 1 2125.68 ⫺205.95 ⫺ v(1) ⫺ u(1)(0) d 1 0 ⫺205.95 5.4168

Subject to ⫺145.98d1 ⫹ 19.148d2 ⫽ ⫺13.318 d1 ⫹ d2 ⭓ 0 The subproblem solution is d (1) ⫽ (⫹0.10434, ⫺0.10434)T, and the multipliers are v(2) ⫽ ⫺0.02497, u(2) ⫽ 0.38822. At the new point x(2) ⫽ (0.40183, 0.59817)T, the objective function value is 0.24036 and the constraint value is 2.7352. The results of the next iteration, x(3) ⫽ (0.74969, 0.25031)T

ƒ(x(3)) ⫽ 0.18766

h(x(3)) ⫽ 13.416

indicate that the constraint violation has increased considerably while the objective function value has decreased somewhat. Comparing the status at x(1) and x(3), it is evident that the iterations show no real improvement. In fact, both the objective function value and the equality constraint violation have increased in proceeding from x(1) to x(3). The solution to the problem of unsatisfactory convergence is, as in the unconstrained case, to perform a line search from the previous solution estimate in the direction obtained from the current QP subproblem solution. However, since in the constrained case both objective function improvement and reduction of the constraint infeasibilities need to be taken into account, the line search must be carried using some type of penalty function. For instance, as in the case of the SLP strategy advanced by Palacios-Gomez (recall Section 8.1.2), an exterior penalty function of the form

再冘 K

P(x, R) ⫽ ƒ(x) ⫹ R

k⫽1

冘 [min(0, g (x))] 冎 J

[hk (x)]2 ⫹

2

j

j⫽1

could be used along with some strategy for adjusting the penalty parameter R. This approach is illustrated in the next example. Example 10.5

Consider the application of the penalty function line search to the problem of Example 10.4 beginning with the point x(2) and the direction vector d (2) ⫽ (0.34786, ⫺0.34786)T which was previously used to compute the point x(3) directly. Suppose we use the penalty function

10.2

463


P(x, R) ⫽ ƒ(x) ⫹ 10{h(x)2 ⫹ [min(0, g(x)]2} and minimize it along the line x ⫽ x(2) ⫹ ␣d (2) ⫽

冉

冊冉

冊

0.40183 0.34786 ⫹ ␣ 0.59817 ⫺0.34786

Note that at ␣ ⫽ 0, P ⫽ 75.05, while at ␣ ⫽ 1, P ⫽ 1800.0. Therefore, a minimum ought to be found in the range 0 ⭐ ␣ ⭐ 1. Using any convenient line search method, the approximate minimum value P ⫽ 68.11 can be found with ␣ ⫽ 0.1. The resulting point will be x(3) ⫽ (0.43662, 0.56338)T with ƒ(x(3)) ⫽ 0.24682 and h(x(3)) ⫽ 2.6053. To continue the iterations, updated estimates of the multipliers are required. Since d ⫽ x(3) ⫺ x(2) is no longer the optimum solution of the previous subproblem, this value of d cannot be used to estimate the multipliers. The only available updated multiplier values are those associated with d (2), namely v(3) ⫽ ⫺0.005382 and u(3) ⫽ 0.37291. The results of the next four iterations obtained using line searches of the penalty function after each subproblem solution are shown in Table 10.1. As is evident from the table, the use of the line search is successful in forcing convergence to the optimum from poor initial estimates. The use of the quadratic approximation to the Lagrangian was proposed by Wilson [1]. Although the idea was pursued by Beale [2] and by Bard and Greestadt [3], it has not been widely adopted in its direct form. As with Newton’s method, the barriers to the adoption of this approach in engineering applications have been twofold: first, the need to provide second derivative values for all model functions and, second, the sensitivity of the method to poor initial solution estimates. Far from the optimal solution, the function second derivatives and especially the multiplier estimates will really have little

Table 10.1

Results for Example 10.5

Iteration

x1

x2

ƒ

h

v

3 4 5 6 7

0.43662 0.48569 1.07687 0.96637 0.99752 1.0

0.56338 0.56825 1.8666 1.8652 1.99503 2.0

0.24682 0.27599 2.0101 1.8025 1.9901 2.0

2.6055 2.5372 0.07108 0.10589 0.00498 0.0

⫺0.005382 ⫺0.3584 ⫺0.8044 ⫺1.6435 ⫺1.9755 ⫺2.0

⬁

464


relevance to defining a good search direction. (For instance, Table 10.1, v(3) ⫽ ⫺5.38 ⫻ 10⫺3, while v* ⫽ ⫺2.) Thus, during the initial block of iterations, the considerable computational burden of evaluating all second derivatives may be entirely wasted. A further untidy feature of the above algorithm involves the strategies required to adjust the penalty parameter of the line search penalty function. First, a good initial estimate of the parameter must somehow be supplied; second, to guarantee convergence, the penalty parameter must in principle be increased to large values. The manner in which this is done requires heuristics of the type considered in Chapter 6. Recent work discussed in the next section has largely overcome these various shortcomings of the quadratic approximation approach.

10.3 VARIABLE METRIC METHODS FOR CONSTRAINED OPTIMIZATION

In Chapter 3 we noted that in the unconstrained case the desirable improved convergence rate of Newton’s method could be approached by using suitable update formulas to approximate the matrix of second derivatives. Thus, with the wisdom of hindsight, it is not surprising that, as first shown by GarciaPalomares and Mangasarian [4], similar constructions can be applied to approximate the quadratic portion of our Lagrangian subproblems. The idea of approximating ⵜ2L using quasi-Newton update formulas that only require differences of gradients of the Lagrangian function was further developed by Han [5, 6] and Powell [7, 8]. The basic variable metric strategy proceeds as follows. Constrained Variable Metric Method

Given initial estimates x 0, u0, v0 and a symmetric positive-definite matrix H0. Step 1. Solve the problem Minimize

ⵜƒ(x(t))Td ⫹ –12 d TH(t)d

Subject to hk (x(t)) ⫹ ⵜhk (x(t))Td ⫽ 0

k ⫽ 1, . . . , K

gj(x ) ⫹ ⵜgj(x ) d ⭓ 0

j ⫽ 1, . . . , J

(t)

(t) T

Step 2. Select the step size ␣ along d (t) and set x(t⫹1) ⫽ x(t) ⫹ ␣d (t). Step 3. Check for convergence. Step 4. Update H(t) using the gradient difference ⵜx L(x(t⫹1), u(t⫹1), v(t⫹1)) ⫺ ⵜx L(x(t), u(t⫹1), v(t⫹1))

in such a way that H(t⫹1) remains positive definite.

10.3

VARIABLE METRIC METHODS FOR CONSTRAINED OPTIMIZATION

465

The key choices in the above procedure involve the update formula for H(t) and the manner of selecting ␣. Han [4, 5] considered the use of several wellknown update formulas, particularly DFP. He also showed [4] that if the initial point is sufficiently close, then convergence will be achieved at a superlinear rate without a step-size procedure or line search by setting ␣ ⫽ 1. However, to assure convergence from arbitrary points, a line search is required. Specifically, Han [5] recommends the use of the penalty function

再冘兩h (x)兩 ⫺ 冘 min[0, g (x)]冎 K

J

P(x, R) ⫽ ƒ(x) ⫹ R

k

j

k⫽1

j⫽1

to select ␣* so that P(x(␣*)) ⫽ min P(x(t) ⫹ ␣ d (t), R) 0⭐␣⭐␦

where R and ␦ are suitably selected positive numbers. Powell [7], on the other hand, suggests the use of the BFGS formula together with a conservative check that ensures that H(t) remains positive definite. Thus, if z ⫽ x(t⫹1) ⫺ x(t) and y ⫽ ⵜx L(x(t⫹1), u(t⫹1), v(t⫹1)) ⫺ ⵜx L(x(t), u(t⫹1)v(t⫹1)) Then define

␪⫽

冦

1

if zTy ⭓ 0.2zTH(t)z

0.8zTH(t)z zTH(t)z ⫺ zTy

otherwise

(10.9)

and calculate w ⫽ ␪y ⫹ (1 ⫺ ␪)H(t)z

(10.10)

Finally, this value of w is used in the BFGS updating formula, H(t⫹1) ⫽ H(t) ⫺

H(t)zzTH(t) wwT ⫹ T zTH(t)zT zw

(10.11)

Note that the numerical value 0.2 is selected empirically and that the normal BFGS update is usually stated in terms of y rather than w.

466


On the basis of empirical testing, Powell [8] proposed that the step-size procedure be carried out using the penalty function

冘 ␮ 兩h (x)兩 ⫺ 冘 ␴ min(0, g (x)) K

P(x, ␮, ␴) ⫽ ƒ(x) ⫹

J

k

k

j

k⫽1

j

(10.12)

j⫽1

where for the first iteration ␮k ⫽ 兩vk兩

␴j ⫽ 兩uj兩

and for all subsequent iterations t (t) 1 (t⫺1) ␮(t) ⫹ 兩v(t) k ⫽ max{兩vk 兩, – 2 (␮k k 兩)}

(10.13)

(t) 1 (t⫹1) ␴ (t) ⫹ 兩u(t) j ⫽ max{兩uj 兩, – 2 (␴ j j 兩)}

(10.14)

The line search could be carried out by selecting the largest value of ␣, 0 ⭐ ␣ ⭐ 1, such that P(x(␣)) ⬍ P(x(0))

(10.15)

However, Powell [8] prefers the use of quadratic interpolation to generate a sequence of values of ␣k until the more conservative condition dP P(x(␣k)) ⭐ P(x(0)) ⫹ 0.1␣k (x(0)) d␣

(10.16)

is met. It is interesting to note, however, that examples have been found for which the use of Powell’s heuristics can lead to failure to converge [9]. Further refinements of the step-size procedure have been reported [10], but these details are beyond the scope of the present treatment. We illustrate the use of a variant of the constrained variable metric (CVM) method using update (10.11), penalty function (10.12), and a simple quadratic interpolation-based step-size procedure. Example 10.6

Solve Example 10.1 using the CVM method with initial metric H0 ⫽ I. At the initial point (2, 1), the function gradients are 23 47 T ⵜƒ ⫽ (–– 4 , ⫺–– 4 )

ⵜh ⫽ (1, 2)T

Therefore, the first subproblem will take the form

ⵜg ⫽ (1, 1)T

10.3


Minimize

23 47 1 T (–– 4 , ⫺–– 4 )d ⫹ – 2 d Id

Subject to

(1, 2)d

467

⫽0

(1, 1)d ⫹ 2 ⭓ 0 It is easy to show that the problem solution lies at the intersection of the two constraints. Thus, d 0 ⫽ (⫺4, 2)T, and the multipliers at this point are solutions of the system

再冉冊冉冊冎冉冊冉冊 23 –– 4 –– ⫺ 47 4

⫹

⫺4

⫽v

2

1 1 ⫹u 2 1

46 (1) 53 or v(1) ⫽ ⫺ –– ⫽ –– 4 , u 4 . For the first iteration, we use the penalty parameters

46 ␮(1) ⫽ 兩⫺ –– 4 兩

and

53 ␴ (1) ⫽ 兩–– 4 兩

The penalty function (10.12) thus take the form ⫺2 46 53 P ⫽ 6x1x⫺1 2 ⫹ x2x1 ⫹ –– 4 兩x1x2 ⫺ 2兩 ⫺ –– 4 min(0, x1 ⫹ x2 ⫺ 1)

We now conduct a one-parameter search of P on the line x ⫽ (2, 1)T ⫹ ␣(⫺4, 2)T. At ␣ ⫽ 0, P(0) ⫽ 12.25. Suppose we conduct a bracketing search with ⌬ ⫽ 0.1. Then, P(0 ⫹ 0.1) ⫽ 9.38875

and

P(0.1 ⫹ 2(0.1)) ⫽ 13.78

Clearly, the minimum on the line has been bounded. Using quadratic interpolation on the three trial points of ␣ ⫽ 0, 0.1, 0.3, we obtain ␣ ⫽ 0.1348 with P(␣) ⫽ 9.1702. Since this is a reasonable improvement over P(0), the search is terminated with this value of ␣. The new point is x(1) ⫽ (2, 1)T ⫹ (0.1348)(⫺4, 2) ⫽ (1.46051, 1.26974) We now must proceed to update the matrix H. Following Powell, we calculate z ⫽ x(1) ⫺ x 0 ⫽ (⫺0.53949, 0.26974)T Then

468


冢冣

冉冊冉冊冉冊冉冊冉冊

23 46 4 ⵜx L(x 0, u(1), v(1)) ⫽ 47 ⫺ ⫺ 4 ⫺ 4 ⵜx L(x(1), u(1), v(1)) ⫽

冉冉

1

⫺

2

冊冉冊冉

3.91022

⫺

⫺4.96650

46 4

1

53 4

4

⫽

1

冊冉冊

1.26974

⫺

⫺2

⫺

1.46051

53 1 4 1

冊

5.26228

⫽

4.81563

Note that both gradients are calculated using the same multiplier values u(1), v(1). By definition, y ⫽ ⵜx L(x(1)) ⫺ ⵜx L(x 0) ⫽ (1.26228, 6.81563)T Next, we check condition (10.9): zTy ⫽ 1.15749 ⬎ 0.2zTIz ⫽ 0.2(0.3638) Therefore, ␪ ⫽ 1 and w ⫽ y. Using (10.11), the update H(1) is H(1) ⫽ I ⫺ ⫽

zzT yyT ⫹ T 2 㛳z㛳 zy

冉冊

冉

⫹ (1.15749)⫺1 ⫽

冉

冊

1 0 ⫺0.53949 ⫺ (0.3638)⫺1 (⫺0.53949, 0.26974) 0 1 ⫹0.26974

冉

冊

1.26228 (1.26228, 1.26228) 1.26228

冊

1.57656 7.83267 7.83267 40.9324

Note that H(1) is positive definite. This completes on iteration. We will carry out the second in abbreviated form only. The subproblem at x(1) is

冉

(3.91022, ⫺4.96650)d ⫹ –12 d T

Subject to

(1.26974, 1.46051)d ⫺ 0.14552 ⫽ 0 d1 ⫹ d2 ⫹ 1.73026 ⭓ 0

The solution of the quadratic program is

冊

1.57656 7.83267 d 7.83267 40.9324

Minimize

10.3

469


d (1) ⫽ (⫺0.28911, 0.35098)T At this solution, the inequality is loose, and hence u(2) ⫽ 0. The other multiplier value is v(2) ⫽ 4.8857. The penalty function multipliers are updated using (10.13) and (10.14):

冉冉

␮(2) ⫽ max 兩4.8857兩, ␴ (2) ⫽ max 兩0兩,

冊

53 –– 4 ⫹ 0

2

冊

46 4.8857 ⫹ ⫽ 8.19284 4 2 ⫽ 6.625

The penalty function now becomes P(x(␣)) ⫽ ƒ(x) ⫹ 8.19284兩x1x2 ⫺ 2兩 ⫺ 6.625 min(0, x1 ⫹ x2 ⫺1) where x(␣) ⫽ x(1) ⫹ ␣d (1)

0⭐␣⭐1

At ␣ ⫽ 0, P(0) ⫽ 8.68896, and the minimum occurs at ␣ ⫽ 1, P(1) ⫽ 6.34906. The new point is x(2) ⫽ (1.17141, 1.62073) with ƒ(x(2)) ⫽ 5.5177

h(x(2)) ⫽ ⫺0.10147

and

The iterations continue with an update of H(1). The details will not be elaborated since they are repetitious. The results of the next four iterations are summarized below. Iteration

x1

x2

ƒ

h

v

3 4 5 6

1.14602 1.04158 0.99886 1.00007

1.74619 1.90479 1.99828 1.99986

5.2674 5.03668 5.00200 5.00000

0.001271 ⫺0.01603 ⫺0.003994 ⫺1.9 ⫻ 10⫺6

⫺0.13036 ⫺0.17090 ⫺0.45151 ⫺0.50128

Recall that in Example 10.3, in which analytical second derivatives were used to formulate the QP subproblem, comparable solution accuracy was attained in four iterations. Thus, the quasi-Newton result obtained using only first derivatives is quite satisfactory, especially in view of the fact that the line searches were all carried out only approximately.

470


It should be reemphasized that the available convergence results (superlinear rate) [6, 11] assume that the penalty function parameters remain unchanged and that exact line searches are used. Powell’s modifications (10.13) and (10.14) and the use of approximate searches thus amount to useful heuristics justified solely by numerical experimentation. Finally, it is noteworthy that an alternative formulation of the QP subproblem has been reported by Biggs as early as 1972 [12]. The primary differences of that approach lie in the use of an active constraint strategy to select the inequality constraints that are linearized and the fact that the quadratic approximation appearing in the subproblem is that of a penalty function. In view of the overall similarity of that approach to the Lagrangian-based construction, we offer no elaboration here, but instead invite the interested reader to study the recent exposition of this approach offered in reference 13 and the references cited therein. It is of interest that as reported by BartholomewBiggs [13], a code implementing this approach (OPRQ) has been quite successful in solving a number of practical problems, including one with as many as 79 variables and 190 constraints. This gives further support to the still sparse but growing body of empirical evidence suggesting the power of CVM approaches. 10.4

DISCUSSION

Having outlined the general form of the family of CVM algorithm, we continue in this section to discuss some practical implementation features as well as to review the relationship between GRG and CVM methods. 10.4.1

Problem Scaling

First of all, as noted by Powell [7] and underscored by the experience of Berna et al. [14], the variable metric approach should be and apparently is insensitive to scaling of the constraints. However, although the method should similarly be insensitive to variable scaling, the experience reported in reference 14 indicates that scaling is in fact important. Berna et al. recommend scaling all variables so that they take on values between 0.1 and 10. 10.4.2

Constraint Inconsistency

As can be seen from Example 10.6 and is true in general, this family of methods will generate a series of infeasible points. This can lead to various difficulties ranging from model evaluation failures, because the model functions may be invalid for some variable values, to the generation of linearized constraint sets that are inconsistent. The former difficulty is a common feature of all exterior point methods and can be remedied in part by carefully formulating the model to avoid the possibility of obvious sources of numerical problems, such as nonpositive arguments of logarithms or power functions or

10.4

DISCUSSION

471

division by very small numbers or zero. The situation of inconsistent linearized constraint sets can in principle occur even though the original problem has a nonempty feasible region, particularly for very infeasible initial solution estimates. The device recommended by Powell [7] is to insert a dummy variable ␰, constrained to lie in the range 0 ⭐ ␰ ⭐ 1, into the subproblem formulation. All linearized equality constraints are written ⵜhk (x)d ⫹ ␰hk (x) ⫽ 0

and the linearizations of all inequalities with negative values are written ⵜgj(x)d ⫹ ␰gj(x) ⭓ 0

The QP objective function is modified by adding the term (1 ⫺ ␰)M, where M is a large positive constant. When ␰ ⫽ 1, the original subproblem is obtained. If no feasible d can be found, then a value of ␰ ⬍ 1 will allow additional freedom to find a d satisfying the modified subproblem constraints. The use of this device is strongly endorsed by Berna et al. [14] for dealing with large-scale applications. 10.4.3

Modification of H(t)

As noted by Chao et al. [15] in the context of structural optimization problems, two additional numerical difficulties may arise in the use of the CVM algorithm: excessively large direction vector values d and nearly singular values of H(t). The former may occur when at the optimal solution of the original problem the total number of active constraints (inequalities plus equalities) is less than the number of variables. The latter situation can occur because, even though the updating procedure does maintain H(t) symmetric and positive definite, the matrix may become nearly positive semidefinite. The correction suggested in both instances is to add an identity matrix weighted by an adjustable nonnegative parameter ␲ to the approximation H(t). No systematic adjustment procedure is reported other than a heuristic adjustment based on the resulting magnitude of d. The excellent results obtained using the resulting modified CVM algorithm [15] with several large structural optimization problems when compared to other NLP approaches suggest that such heuristic adjustment can be satisfactory. Nonetheless, this issue merits further investigation, since in principle such modification of the H(t) matrix could cause deterioration of the desirable CVM algorithm convergence properties. 10.4.4

Comparison of GRG with CVM

Since both the accelerated GRG method and the CVM method make use of quasi-Newton updates to improve the search directions deduced from linear approximations, it is reasonable to question whether the two approaches might

472


not in fact be largely equivalent. In the following discussion we show via qualitative arguments that there are substantive differences between the two approaches. For purposes of this discussion we consider only the equalityconstrained case. The inclusion of inequality constraints is merely a matter of detail that the reader might consider as an interesting exercise. Given an equality-constrained problem, both methods, in principle, seek a point that satisfies the Lagrangian necessary conditions r(x, v) ⬅ ⵜƒ(x) ⫺ vT ⵜh(x) ⫽ 0 h(x) ⫽ 0

(10.17) (10.18)

where for simplicity we denote by h the vector of equality constraint functions. The CVM approach can be viewed as an attempt to solve the above set of nonlinear equations by using Newton’s method [16]. Thus, at some initial estimate (x 0, v0), the application of Newton’s method requires the solution of the linearized equations. r(x 0, v0) ⫹ ⵜxr(x 0, v0)T ⌬ x ⫹ ⵜvr(x 0, v0)T ⌬v ⫽ 0 h(x 0) ⫹ ⵜh(x 0)T ⌬ x ⫽ 0 If we substitute for r its definition in terms of the gradient of the Lagrangian, the result is ⵜƒ(x 0)T ⫺ vT ⵜh(x 0) ⫹ d T ⵜ2x L(x 0, v0) ⫽ 0

(10.19)

h(x 0) ⫹ ⵜh(x 0)Td ⫽ 0

(10.20)

As shown in Section 10.3, these equations are simply the necessary conditions for a minimum of the problem Minimize

ⵜƒ(x 0)Td ⫹ –12 d T ⵜ2x L(x 0, v0)d

Subject to h(x 0) ⫹ ⵜh(x 0)Td ⫽ 0 Evidently, the CVM strategy can thus be viewed as a process of solving Eqs. (10.19) and (10.20) by using a quasi-Newton method to approximate ⵜ2x L. Notice that in general the CVM method does not require x 0 to satisfy h(x 0) ⫽ 0. Since neither the condition h(x) ⫽ 0 nor the requirement min ƒ(x) is directly enforced in solving these equations, it proves necessary to execute a line search along d to minimize a penalty function. Note that in Eq. (10.19) the term

10.4

DISCUSSION

473

ⵜƒ(x 0)T ⫺ vT ⵜh(x 0)

includes the new value of the multiplier v. Thus, the need for updating the approximation H of ⵜ2x L by using the difference [ⵜƒ(x(1))T ⫺ (v(1))T ⵜh(x(1))] ⫺ [ⵜƒ(x 0)T ⫺ (v(1))T ⵜh(x 0)] is apparent. On the other hand, the accelerated GRG method approaches the solution of Eqs. (10.17) and (10.18) by first partitioning the variables and function gradients into x ⫽ (ˆx, x)

ⵜh ⫽ (J, C)

ˆ ⵜƒ) ⵜƒ ⫽ (ⵜƒ,

to obtain the system ˆ T ⫺ vTJ(x) ⫽ 0 ⵜƒ(x)

(10.21)

ⵜƒ(x) ⫺ v C(x) ⫽ 0

(10.22)

T

T

h(x) ⫽ 0 Then condition (10.21) is forced by defining ˆ TJ⫺1(x) vT ⫽ ⵜƒ(x)

(10.23)

From the implicit function theorem of calculus, given a system of equations h(x) ⫽ 0, if J(x) is nonsingular in the neighborhood of a point x 0, then there exists a set of functions q such that xˆ ⫽ q(x) Consequently, the dependence on x of all functions in (10.22) and (10.23) can be reduced to a dependence on x only. Condition (10.22) can thus be restated to require that ˜ ⵜƒ(x) ⫽0

(10.24)

Again the application of Newton’s method at a point x 0 that is feasible, that is, where h(x 0) ⫽ 0, results in the system ˜ 0)T ⫹ d T ⵜ2ƒ(x ˜ 0) ⫽ 0 ⵜƒ(x

or

˜ 0)T(ⵜ2ƒ(x ˆ 0))⫺1 d T ⫽ ⫺ ⵜƒ(x

(10.25)

where ⵜ2ƒ˜ is the ‘‘reduced’’ second derivative. An exact expression for this quantity can be derived in terms of the first and second derivatives of the

474


problem functions as well as v0 as given by Eq. (10.23) [17]. For present purposes, that expression is not really essential, since ⵜ2ƒ˜ is approximated by a quasi-Newton matrix H that is updated by employing the reduced gradient differences ˜ (1)) ⫺ ⵜƒ(x ˜ 0) ⵜƒ(x

Note that, implicitly, ⵜƒ˜ (x(1)) uses v(1) and ⵜƒ˜ (x 0) employs v0, calculated at x(1) and x 0, respectively, using Eq. (10.23). In the GRG method the correction dˆ is calculated so that constraint feasibility is maintained. Therefore, the line search on d can be executed using ƒ(x) directly. The conceptual differences between the GRG and CVM strategies are now quite evident. First, the multipliers for GRG are estimated via (10.23) using only first-derivative information, whereas v in the CVM approach satisfies (10.19) and thus does reflect second-order information to some degree. Second, the size of the quasi-Newton approximation matrix H is quite different in the two cases: N ⫻ N for CVM and (N ⫺ K) ⫻ (N ⫺ K) in the case of GRG. Third, GRG executes its line search by using the objective function directly, while CVM employs a penalty function. Finally, GRG repeatedly solves the constraint equations to maintain feasibility, while CVM merely deals with the constraint linearizations. While the effects of the differences in the multiplier estimates are not obvious, it is clear that for large problems the differences in the size of the matrix H that must be stored and updated can be quite significant. It is interesting to note that Berna et al. [14] have shown that the implicit variable elimination strategy used in GRG can be employed within the CVM constructions to result in a QP subproblem of dimension N ⫺ K. However, in their scheme the full H matrix is always updated and then reduced to size (N ⫺ K) ⫻ (N ⫺ K) by implicit variable elimination calculations. While the storage and updating of H represents a significant computational burden, quite clearly the major difference between GRG and CVM lies in the iterations to maintain constraint feasibility. On balance, the above considerations clearly suggest that if the NLP model to be solved is sufficiently stable to allow substantial excursions outside the feasible region, then the CVM approach is to be preferred. However, if the model cannot be evaluated at points that are infeasible, then the accelerated GRG method is best employed. An interesting example of this situation is given in the study reported by Biegler and Hughes [18], who employed the CVM approach to optimize the linearized response surface models generated from rigorous chemical process simulation models. They found that substantial constraint violations could not be tolerated, and hence the CVM approach had to be modified to include iterations to satisfy constraints. Evidently, in that work an accelerated GRG approach could well have been used, avoiding the need to update the complete H matrix. Further discussion of issues arising in the adaptation of the CVM method to larger scale problems is given by

10.5

SUMMARY

475

Gill et al. [19]. It is clear from the above discussion that both approaches have an important role in engineering optimization and constitute the two best approaches to solving constrained nonlinear optimization problems presently available. 10.5

SUMMARY

In this chapter we have studied the consequences of attempting to use higher order, specifically quadratic, approximations to nonlinear problem functions. We first examined full quadratic approximation and found that a problem with quadratic constraints was no easier to solve than one with general nonlinear constraints. Next we considered the formulation of QP subproblems: quadratic objective functions and linear constraints. We found that such subproblems were convenient to solve, but because they involve only the original constraints in linearized form, the result is not a significant improvement over the SLP approaches of Chapter 8. We then examined the use of a subproblem objective function whose quadratic term was the second derivative of the Lagrangian function. We found that if a solution of the subproblem satisfies the necessary conditions and the second-order sufficient conditions for a local minimum, then the corresponding point will also satisfy the sufficient conditions for the original problem. Moreover, the Lagrangian multipliers of a given subproblem will conveniently serve as multipliers for the construction of the next subproblem. Thus, the basis was established for an efficient algorithm for generating good search directions. However, we also observed that the method does not necessarily generate feasible directions, and hence to ensure convergence, we found it expedient to conduct a line search along the generated direction using an exterior penalty function. The remaining major difficulty was the need to provide second derivatives of all problem functions. This difficulty was resolved by using quasi-Newton methods, which only require differences of gradients of the Lagrangian function to approximate and update the second-derivative term. The result was a CVM algorithm that can be shown to possess a local superlinear convergence rate even without line searching. Of course, for global convergence, line searching using a penalty function proved necessary. Our discussion of successive QP methods proceeded with a brief summary of enhancements that have been suggested by various investigators to treat cases of subproblem constraint inconsistency and ill-conditioning or near singularity of the quasiNewton matrix. Finally, we compared the relative merits of the GRG algorithm with quasi-Newton acceleration and the CVM algorithm. We found that the key differences are that the CVM algorithm generates infeasible intermediate points while GRG maintains feasibility and that GRG uses first-order multiplier estimates while CVM uses some level of higher order information. Of course, from a storage viewpoint, the quasi-Newton matrix of the GRG

476


method is, in general, smaller than that of the CVM algorithm. On balance, we found that both methods have features that are unique and hence both are important to engineering optimization. REFERENCES 1. Wilson, R. B., ‘‘A Simplicial Algorithm for Concave Programming,’’ Ph.D. dissertation, Harvard University, Cambridge, MA, 1963. 2. Beale, E. M. L., ‘‘Numerical Methods,’’ in Nonlinear Programming (J. Abadie, Ed.), North-Holland, Amsterdam, 1967. 3. Bard, Y., and J. L. Greenstadt, ‘‘A Modified Newton Method for Optimization with Equality Constraints,’’ in Optimization (R. Fletcher, Ed.), Academic, New York, 1969. 4. Garcia-Palomares, U. M., and O. L. Mangasarian, ‘‘Superlinearly Convergent Quasi-Newton Algorithms for Nonlinearly Constrained Optimization Problem,’’ Math. Prog., 11, 1–13 (1976). 5. Han, S. P., ‘‘Superlinearly Convergent Variable Metric Algorithms for General Nonlinear Programming Problems,’’ Math. Prog., 11, 263–282 (1976). 6. Han, S. P., ‘‘A Globally Convergent Method for Nonlinear Programming,’’ J. Opt. Theory Appl. 22, 297–309 (1977). 7. Powell, M. J. D., ‘‘A Fast Algorithm for Nonlinearly Constrained Optimization Calculations,’’ in Numerical Analysis, Dundee 1977 (G. A. Watson, Ed.), Lecture Notes in Mathematics No. 630, Springer-Verlag, New York, 1978. 8. Powell, M. J. D., ‘‘Algorithms for Nonlinear Functions that Use Lagrangian Functions,’’ Math. Prog., 14, 224–248 (1978). 9. Chamberlain, R. M., ‘‘Some Examples of Cycling in Variable Metric Methods for Constrained Minimization,’’ Math. Prog., 16, 378–383 (1979). 10. Mayne, D. Q., ‘‘On the Use of Exact Penalty Functions to Determine Step Length in Optimization Algorithms,’’ in Numerical Analysis, Dundee, 1979 (G. A. Watson, Ed.), Lecture Notes in Mathematics No. 773, Springer-Verlag, New York, 1980. 11. Powell, M. J. D., ‘‘The Convergence of Variable Metric Methods for Nonlinearly Constrained Optimization Calculations,’’ in Nonlinear Programming, Vol. 3 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson Eds.), Academic, New York, 1978. 12. Biggs, M. C., ‘‘Minimization Using Recursive Equality Quadratic Programming,’’ in Numerical Methods for Nonlinear Optimization (F. A. Lootsma, Ed.), Academic, New York, 1972. 13. Bartholemew-Biggs, M. C., ‘‘Recursive Quadratic Programming Based on Penalty Functions for Constrained Minimization,’’ in Nonlinear Optimization: Theory and Algorithms (L. C. W. Dixon, E. Spedicato, and G. P. Szego, Eds.), Birkhauser, Boston, 1980. 14. Berna, T. J., M. H. Locke, and A. W. Westerberg, ‘‘A New Approach to Optimization of Chemical Processes,’’ AICHE J., 26, 37–43 (1980). 15. Chao, N. H., S. J. Fenvers, and A. W. Westerberg, ‘‘Application of a Reduced Quadratic Programming Technique to Structural Design,’’ in Proc. Intern. Symp. Optimum Structural Design, Tucson, AZ, Oct. 1981, pp. 8.5–8.12.

PROBLEMS

477

16. Powell, M. J. D., ‘‘Constrained Optimization Using Variable Metric Methods,’’ in Nonlinear Optimization: Theory and Algorithms (L. C. W. Dixon, E. Spedicato, and G. P. Szego, Eds.), Birkhauser, Boston, 1980. 17. Reklaitis, G. V., and D. J. Wilde, ‘‘A Computationally Compact Sufficient Condition for Constrained Minima,’’ Oper. Res., 17, 425–436 (1969). 18. Biegler, L. T., and R. R. Hughes, ‘‘Approximation Programming of Chemical Process with Q / LAP,’’ Chem. Eng. Prog., 77, 76–83 (1981). 19. Gill, P. E., W. Murray, M. A. Saunders, and M. H. Wright, ‘‘QP-Based Method for Large Scale Nonlinearly Constrained Optimization,’’ in Nonlinear Programming, Vol. 4 (O. L. Mangasarian, R. R. Meyer, and S. M. Robinson, Eds.), Academic, New York, 1981.

PROBLEMS

10.1. What is the major disadvantage associated with a solution technique based upon direct use of full quadratic approximations to all functions in the nonlinear program? 10.2. Consider the general NLP problem of Eqs. (10.7), (10.8a), and (10.8b). Show that if d ⫽ 0 satisfies the second-order sufficient conditions for a local minimum, then the approximating point satisfies the sufficient conditions for the original problem. 10.3. Outline an implementation of a successive Lagrangian QP algorithm that would employ the more conservative step adjustment strategy of the Griffith and Stewart SLP algorithm. Discuss the advantages and disadvantages relative to the penalty function strategy. 10.4. It is possible to modify the step-size adjustment mechanism of the CVM method so as to generate feasible points as in the GRG method. Explain how this might be accomplished. 10.5. Compare the treatment of inequality constraints in the GRG and CVM algorithms. How do the methods of estimating multiplier values differ? 10.6. Suppose the CVM algorithm were employed with a problem involving a quadratic objective function and quadratic inequality constraints. How many iterations are likely to be required to solve the problem, assuming exact arithmetic? What assumptions about the problem are likely to be necessary in making this estimate? 10.7. Construct a full quadratic approximation to the problem Minimize

ƒ(x) ⫽ x1 ⫹ x22 ⫹ x33

Subject to g1(x) ⫽ 1 ⫺ 64x1⫺2x2⫺4x3⫺6 ⭓ 0 all xi ⭓ 0 at the point (–12 , –12 , –12 ). Is the resulting problem a convex problem?

478


10.8. Consider the NLP Minimize

⫺1 ƒ(x) ⫽ x⫺1 1 ⫹ x2

Subject to h(x) ⫽ –21 x12 ⫹ x22 ⫺ 1 ⫽ 0 x1, x2 ⭓ 0 (a) Construct a full quadratic approximation to the problem at the point x 0 ⫽ (–34 , –34 ). (b) Solve the resulting subproblem for the correction vector d and set x1 ⫽ x 0 ⫹ d. (c) Is the resulting point an improvement over x 0? How might it be further improved? 10.9. Suppose the NLP problem of problem 10.8 is approximated at the same x 0 by the Lagrangian-based QP subproblem [Eqs. (10.3 and (10.4)]. (a) Solve the resulting subproblem for the correction vector d and set x1 ⫽ x 0 ⫹ d. (b) Calculate the new estimates of the multipliers. (c) Formulate the next QP subproblem. 10.10. Consider the problem Minimize

ƒ(x) ⫽ –13 (x1 ⫹ 3)3 ⫹ x22

Subject to h(x) ⫽ x32 ⫺ x1 ⫹ 1 ⫽ 0 x1 ⭓ 1 (a) Given the point x 0 ⫽ (2, 1), construct the initial subproblem for the CVM algorithm. (b) Solve the subproblem to obtain d. (c) Calculate the step length using Powell’s procedure and calculate the next point. (d) Construct the next subproblem of the CVM method. 10.11. Given the problem Minimize

ƒ(x) ⫽ 3x21 ⫺ 4x2

Subject to h(x) ⫽ 2x1 ⫹ x2 ⫺ 4 ⫽ 0 g(x) ⫽ 37 ⫺ x21 ⫺ x22 ⭓ 0 40 1 the point x 0 ⫽ (⫺1, 6), and the multiplier values (v, u) ⫽ (⫺–– 13 , –– 13 ).

PROBLEMS

479

(a) Formulate the Lagrangian QP subproblem. (b) Show that d ⫽ 0 is the subproblem solution. (c) Show that the point satisfies the second-order conditions for the original problem. 10.12. Suppose that in problem 10.11 the multiplier estimates are (v, u) ⫽ (⫺3, 0.1) at x 0 ⫽ (⫺1, 6). (a) Formulate the Lagrangian QP subproblem. (b) Obtain the subproblem solution. Is it still d ⫽ 0? (c) Calculate the new multiplier values and check the sufficient conditions for a local minimum. 10.13. Consider problem 10.11 with the given point x 0 ⫽ (⫺0.9, 5.8) and the multiplier values (v, u) ⫽ (⫺3, 0.1). (a) Formulate and solve the Lagrangian QP subproblem to obtain the new x and new values of the multipliers. (b) At the new x, calculate the multiplier values as they would be obtained in the GRG method. (c) Compare the results of (a) and (b) to the optimal multiplier values 40 1 (v, u) ⫽ (⫺–– 13 , –– 13 ). 10.14. (a) Carry out three iterations using the Powell variant of the CVM algorithm for solution of the problem Minimize

ƒ(x) ⫽ log(1 ⫹ x21) ⫺ x2

Subject to h(x) ⫽ (1 ⫹ x12)2 ⫹ x22 ⫺ 4 ⫽ 0 using the starting point (1, 1). (b) Verify that the solution is x* ⫽ (0, 兹3). 10.15. The optimal design of a through circulation system for drying catalyst pellets involves the choice of fluid velocity x1 and bed depth x2 so as to maximize the production rate. The resulting problem takes the form [see Thygeson, J. R., and E. D. Grossmann, AICHE J., 16, 749 (1970)] Minimize

ƒ(x) ⫽ 0.0064x1[exp(⫺0.184x 10.3x2) ⫺ 1]

Subject to g1(x) ⫽ 1.2 ⫻ 1013 ⫺ (3000 ⫹ x1)x12x2 ⭓ 0 g2(x) ⫽ 4.1 ⫺ exp(0.184x 0.3 1 x2) ⭓ 0 where x1 and x2 both must be nonnegative. A reasonable starting point is x 0 ⫽ (30,000, 0.25). (a) For solution via the CVM method, is function or variable scaling likely to be useful? Suggest a rescaled form of the problem.

480


(b) At x 0, both constraints are loose. Check whether ⵜ2ƒ is positive definite at x 0. What problems are likely to occur in solving the first subproblem? How are they avoided in the CVM algorithm? (c) Solve the problem using a successive QP or CVM program. [The solution is x* ⫽ (31,766, 0.342).]

11 STRUCTURED PROBLEMS AND ALGORITHMS This chapter is devoted to the study of optimization problems with special structures. So far, we have discussed a number of methods for solving general NLP problems. When the objective function, constraints, and variables have certain forms (e.g., quadratic objective function, linear constraints, integer variables), special-purpose algorithms can be developed to take advantage of these special forms. Often these algorithms are more efficient than the general-purpose algorithms and are capable of solving larger problems. In this chapter, we study the following structured problems and their algorithms: 1. Integer Programming. Specialized LP problems with the added restriction that some or all of the design variables must be integer valued. 2. Quadratic Programming. Optimization problems with a quadratic objective function and linear constraints. 3. Goal Programming. Optimization problems with multiple objective functions that conflict with one another.

11.1

INTEGER PROGRAMMING

An integer linear programming problem, henceforth called an integer program, is an LP problem wherein some or all of the decision variables are restricted to be integer valued. A pure integer program is one in which all the variables are restricted to be integers. A mixed-integer program restricts some of the variables to be integers while others can assume continuous (fractional) values. Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

481

482

STRUCTURED PROBLEMS AND ALGORITHMS

The reason for considering integer programs is that many practical problems require integer solutions. To solve such problems, we could simply solve the linear program, ignoring the integer restrictions, and then either round off or truncate the fractional values of the LP optimal solution to get an integer solution. Of course, while doing this, we have to be careful that the resulting solution stays feasible. Such an approach is frequently used in practice, especially when the values of the variables are so large that rounding or truncating produces negligible change. But in dealing with problems in which the integer variables assume small values, rounding and truncating may produce a solution far from the true optimal integer solution. In addition, for large problems such a procedure can become computationally expensive. For instance, if the optimal LP solution is x1 ⫽ 2.4 and x2 ⫽ 3.5, then we have to try four combinations of integer values of x1 and x2 that are closest to their continuous values, namely, (2, 3) (2, 4), (3, 3), and (3, 4). The one that is feasible and closest to the LP optimal value of the objective function will be an approximate integer solution. With just 10 integer variables, we have to try 210 ⫽ 1024 combinations of integer solutions! Even after examining all such combinations, we cannot guarantee that one of them is an optimal integer solution to the problem.

11.1.1

Formulation of Integer Programming Models

The formulation of integer programs is illustrated by the following examples. Example 11.1

Fixed-Charge Problem

Consider a production planning problem with N products such that the jth product requires a fixed production or setup cost Kj independent of the amount produced and a variable cost Cj per unit proportional to the quantity produced. Assume that every unit of product j requires aij units of resource i and there are M resources. Given that product j, whose sales potential is dj, sells for $pj per unit and no more than bi units of resource i are available (i ⫽ 1, 2, . . . , M), the problem is to determine the optimal product mix that maximizes the net profit. Formulation. The total cost of production (fixed plus variable) is a nonlinear function of the quantity produced. But, with the help of binary (0–1) integer variables, the problem can be formulated as an integer linear program. Let the binary integer variable ␦j denote the decision to produce or not to produce product j. In other words, ␦j ⫽

再

1 if product j is produced 0 otherwise

11.1

INTEGER PROGRAMMING

483

Let xj(⭓ 0) denote the quantity of product j produced. Then the cost of producing xj units of product j is Kj␦j ⫹ Cj xj, where ␦j ⫽ 1 if xj ⬎ 0 and ␦j ⫽ 0 if xj ⫽ 0. Hence, the objective function is

冘 p x ⫺ 冘 (K ␦ ⫹ c x ) N

Maximize Z ⫽

N

j j

j⫽1

j j

j j

j⫽1

The supply constraint for the ith resource is given by

冘a x ⭐b N

ij j

for i ⫽ 1, 2, . . . , M

i

j⫽1

The demand constraint for the jth product is given by xj ⭐ dj␦j xj ⭓ 0

for j ⫽ 1, 2, . . . , N

and ␦j ⫽ 0 or 1

for all j

Note that xj can be positive only when ␦j ⫽ 1, in which case its production is limited by dj and the fixed production cost Kj is included in the objective function. Example 11.2

Handling Nonlinear 0–1 Integer Problems

Consider a nonlinear (binary) integer programming problem: Maximize

Z ⫽ x 21 ⫹ x2 x3 ⫺ x 33

Subject to ⫺2x1 ⫹ 3x2 ⫹ x3 ⭐ 3 x1, x2, x3 僆 {0, 1} This problem can be converted to a linear integer programming problem for solution by using Watters’ transformation [1]. Since, for any positive k and a binary variable xj, x kj ⫽ xj, the objective function immediately reduces to Z ⫽ x1 ⫹ x2 x3 ⫺ x3 Now consider the product term x2 x3. For binary values of x2 and x3, the product x2 x3 is always 0 or 1. Hence, introduce a binary variable y1 such that y1 ⫽ x2 x3. When x2 ⫽ x3 ⫽ 1, we want the value of y1 to be 1, while for all other combinations y1 should be zero. This can be achieved by introducing the following two constraints: x2 ⫹ x3 ⫺ y1 ⭐ 1

⫺x2 ⫺ x3 ⫹ 2y1 ⭐ 0

484


Note when x2 ⫽ x3 ⫽ 1, the above constraints reduce to y1 ⭓ 1 and y1 ⭐ 1, implying y1 ⫽ 1. When x2 ⫽ 0 or x3 ⫽ 0 or both are zero, the second constraint, y1 ⭐

x2 ⫹ x3 2

forces y1 to be zero. Thus the equivalent linear (binary) integer program becomes Maximize

Z ⫽ x1 ⫹ y1 ⫺ x3

Subject to ⫺2x1 ⫹ 3x2 ⫹ x3 ⭐ 3 x2 ⫹ x3 ⫺ y1 ⭐ 1 ⫺x2 ⫺ x3 ⫹ 2y1 ⭐ 0

x1, x2, x3, y1 are (0, 1) variables Remarks 1. Glover and Woolsey [2] have suggested an improvement to Watters’ transformation that introduces a continuous variable rather than an integer variable. 2. The procedure for handling the product of two binary variables can be easily extended to the product of any number of variables. For additional examples of integer programming formulations, the reader is referred to the texts by Ravindran, Phillips, and Solberg [3], and Plane and McMillan [4], Murty [5], and Rardin [6]. Chapter 14 describes a case study on facility location using a mixed-integer programming (MIP) model. 11.1.2

Solution of Integer Programming Problems [3]

Branch-and-Bound Algorithm. The branch-and-bound algorithm is the most

widely used method for solving both pure and MIP problems in practice. Most commercial computer codes for solving integer programs are based on this approach. Basically the branch-and-bound algorithm is just an efficient enumeration procedure for examining all possible integer feasible solutions. We discussed earlier that a practical approach to solving an integer program is to ignore the integer restrictions initially and solve the problem as a linear program. If the LP optimal solution contains fractional values for some integer variables, then by the use of truncation and rounding-off procedures, one can

11.1

INTEGER PROGRAMMING

485

attempt to get an approximate optimal integer solution. For instance, if there are two integer variables x1 and x2 with fractional values 3.5 and 4.4, then we could examine the four possible integer solutions (3, 4), (4, 4), (4, 5), (3, 5) obtained by truncation and rounding methods. We also observe that the true optimal integer solution may not correspond to any of these integer solutions, since it is possible for x1 to have an optimal (integer) value less than 3 or greater than 5. Hence, to obtain the true optimal integer solution, we would have to consider all possible integer values of x1 smaller and larger than 3.5. In other words, the optimal integer solution must satisfy either x1 ⭐ 3

or

x1 ⭓ 4

When a problem contains a large number of integer variables, it is essential to have a systematic method that will examine all possible combinations of integer solutions obtained from the LP optimal solution. The branch-andbound algorithm essentially does this in an efficient manner. To illustrate the basic principles of the branch-and-bound method, consider the following mixed-integer program. Example 11.3

Maximize

Z ⫽ 3x1 ⫹ 2x2

Subject to x1 ⭐ 2 x2 ⭐ 2 x1 ⫹ x2 ⭐ 3.5 x1, x2 ⭓ 0 and integral The initial step is to solve the MIP problem as a linear program by ignoring the integer restrictions on x1 and x2. Call this linear program LP-1. Since we only have two variables, a graphical solution of LP-1 is presented in Figure 11.1. The LP optimal solution is x1 ⫽ 2, x2 ⫽ 1.5, and the maximum value of the objective function Z0 ⫽ 9. Since x2 takes a fractional value, we do not have an optimal solution for the MIP problem. But observe that the optimal integer solution cannot have an objective function value larger than 9, since the imposition of integer restrictions on x2 can only make the LP solution worse. Thus we have an upper bound on the maximum value of Z for the integer program given by the optimal value of LP-1. The next step of the branch-and-bound method is to examine other integer values of x2 that are larger or smaller than 1.5. This is done by adding a new constraint, either x2 ⭐ 1 or x2 ⭓ 2, to the original linear program (LP-1). This creates two new linear programs (LP-2 and LP-3) as follows:

486


Figure 11.1. Solution to LP-1.

LP-2

LP-3

Maximize Z ⫽ 3x1 ⫹ 2x2 Subject to x1 ⭐ 2 x2 ⭐ 2 x1 ⫹ x2 ⭐ 3.5 x2 ⭐ 1 (new constraint) x1, x2 ⭓ 0

Maximize Z ⫽ 3x1 ⫹ 2x2 Subject to x1 ⭐ 2 x2 ⭐ 2 x1 ⫹ x2 ⭐ 3.5 x2 ⭓ 2 (new constraint) x1, x2 ⭓ 0

The feasible regions corresponding to LP-2 and LP-3 are shown graphically in Figures 11.2 and 11.3, respectively. (Note that the feasible region for LP3 is just the straight line AB.) Observe also that the feasible regions of LP-2 and LP-3 satisfy the following: 1. The optimal solution to LP-1 (x1 ⫽ 2, x2 ⫽ 1.5) is infeasible to both LP-2 and LP-3. Thus the old fractional optimal solution will not be repeated. 2. Every integer (feasible) solution to the original (MIP) problem is contained in either LP-2 or LP-3. Thus none of the feasible (integer) solutions to the MIP problem is lost due to the creation of two new linear programs. The optimal solution to LP-2 (Figure 11.2) is x1 ⫽ 2, x2 ⫽ 1, and Z0 ⫽ 8. Thus we have a feasible (integer) solution to the MIP problem. Even though

11.1

INTEGER PROGRAMMING

487


LP-2 may contain other integer solutions, their objective function values cannot be larger than 8. Hence Z0 ⫽ 8 is a lower bound on the maximum value of Z for the mixed-integer program. In other words, the optimal value of Z for the MIP problem cannot be lower than 8. Since we had computed the upper bound earlier as 9, we cannot call the LP-2 solution the optimal integer solution without examining LP-3.


488


The optimal solution to LP-3 (Figure 11.3) is x1 ⫽ 1.5, x2 ⫽ 2, and Z0 ⫽ 8.5. This is not feasible for the mixed-integer program since x1 is taking a fractional value. But the maximum Z value (8.5) is larger than the lower bound (8). Hence it is necessary to examine whether there exists an integer solution in the feasible region of LP-3 whose value of Z is larger than 8. To determine this, we add the constraint either x1 ⭐ 1 or x1 ⭓ 2 to LP-3. This gives two new linear programs LP-4 and LP-5. The feasible region for LP-4 is the straight line DE shown in Figure 11.4, while LP-5 becomes infeasible. The optimal solution to LP-4 (Figure 11.4) is given by x1 ⫽ 1, x2 ⫽ 2, and Z0 ⫽ 7. This implies that every integer solution in the feasible region of LP3 cannot have an objective function value better than 7. Hence the integer solution obtained while solving LP-2, x1 ⫽ 2, x2 ⫽ 1, and Z0 ⫽ 8, is the optimal integer solution to the MIP problem. The sequence of LP problems solved under the branch-and-bound procedure for Example 11.3 may be represented in the form of a network or tree diagram, as shown in Figure 11.5. A network or a tree consists of a set of nodes, which are entry or exit points in the network, and a set of arcs or branches, which are the paths used to enter or exit a node. Node 1 represents the equivalent LP problem (LP-1) of the mixed-integer program ignoring the integer restrictions. From node 1 we branch to node 2 (LP-2) with the help of the integer variable x2 by adding the constraint x2 ⭐ 1 to LP-1. Since we


11.1

INTEGER PROGRAMMING

489

Figure 11.5. Network representation of the branch-and-bound method of Example 11.3.

have an integer optimal solution for node 2, no further branching from node 2 is necessary. Once this type of decision can be made, we say that node 2 has been fathomed. Branching on x2 ⭓ 2 from node 1 results in LP-3 (node 3). Since the optimal solution to LP-3 is fractional, we branch further from node 3 using the integer variable x1. This results in the creation of nodes 4 and 5. Both have been fathomed, since LP-4 has an integer solution while LP-5 is infeasible. The best integer solution obtained at a fathomed node (in this case, node 2) becomes the optimal solution to the MIP problem. Details of the Algorithm

Consider an MIP problem of the form Maximize

Z ⫽ cx

Subject to Ax ⫽ b x⭓0 xj is an integer for j 僆 I where I is the set of all integer variables. The first step is to solve the MIP problem as a linear program by ignoring the integer restrictions. Let us denote by LP-1 the linear program whose optimal value of the objective function is Z1. Assume the optimal solution to LP-1 contains some integer variables at fractional values. Hence we do not have an optimal solution to the MIP problem. But Z1 is an upper bound on the maximum value of Z for the MIP problem.

490


The next step is to partition the feasible region of LP-1 by branching on one of the integer variables at a fractional value. A number of rules have been proposed to select the proper branching variable. They include: 1. Selecting the integer variable with the largest fractional value in the LP solution. 2. Assigning priorities to the integer variables such that we branch on the most important variable first. The importance of an integer variable may be based on one or more of the following criteria: (a) It represents an important decision in the model. (b) Its cost or profit coefficient in the objective function is very large compared to the others. (c) Its value is critical to the model based on the experience of the user. 3. Arbitrary selection rules; for instance, selecting the variable with the lowest index first. Suppose that the integer variable xj is selected for further branching and its fractional value is ␤j in the LP-1 solution. Now we create two new LP problems LP-2 and LP-3 by introducing the constraints xj ⭐ 兩 ␤j and xj ⭓ 兩 ␤j , respectively, where 兩 ␤j is the largest integer less than ␤j, while 兩 ␤j is smallest integer greater than ␤j. (See Figure 11.6.) In other words,

Figure 11.6. Branch-and-bound algorithm.

11.1

INTEGER PROGRAMMING

LP-2

LP-3

Maximize Z ⫽ cx Subject to Ax ⫽ b xj ⭐ 兩 ␤j

Maximize Z ⫽ cx Subject to Ax ⫽ b xj ⭓ 兩 ␤j x⭓0

x⭓0

491

Assume that the optimal solutions to LP-2 and LP-3 are still fractional and hence are infeasible to the MIP problem with integer restrictions. The next step is to select either LP-2 or LP-3 and branch from that by adding a new constraint. Here again a number of rules have been proposed for selecting the proper node (LP problem) to branch from. They include: 1. Using Optimal Value of Objective Function. Considering each of the nodes that can be selected for further branching, we choose the one whose LP optimal value is the largest (for a maximization problem). The rationale for this rule is that the LP feasible region with the largest Z value may contain better integer solutions. For instance, any integer solution obtained by branching from LP-2 cannot have a Z value better than the optimal value of Z for LP-2. 2. Last-In–First-Out Rule. The LP problem that was solved most recently is selected (arbitrarily) for further branching. Once the proper node (LP region) is selected for further branching, we branch out by choosing an integer variable with a fractional value. This process of branching and solving a sequence of linear programs is continued until an integer solution is obtained for one of the linear programs. The value of Z for this integer solution becomes a lower bound on the maximum value of Z for the MIP problem. At this point we can eliminate from consideration all those nodes (LP regions) whose values of Z are not better than the lower bound. We say that these nodes have been fathomed because it is not possible to find a better integer solution from these LP regions than what we have now. As an illustration, consider the tree diagram given in Figure 11.6. With the solution of LP-4 we have a lower bound on the maximum value of Z for the MIP problem given by Z4. In other words, the optimal solution to the MIP problem cannot have a Z value smaller than Z4. Further branching from node 4 is unnecessary, since any subsequent LP solution can only have a Z value less than Z4. In other words, node 4 has been fathomed. Node 5 has also been fathomed since the additional constraints render the LP problem infeasible. This only leaves nodes 6 and 7 for possible branching. Suppose Z6 ⬍ Z4 and Z7 ⬎ Z4. This means node 6 has also been fathomed (implicitly), since none of the integer solutions under node 4 can produce a better value than Z4.

492


However, it is possible for the LP region of node 7 to contain an integer solution better than that of node 4 since Z7 ⬎ Z4. Hence, we select node 7 for further branching and continue. In this manner, an intermediate node (LP problem) is explicitly or implicitly fathomed whenever it satisfies one of the following conditions: 1. The LP optimal solution of that node is integer valued; that is, it is feasible for the MIP problem. 2. The LP problem is infeasible. 3. The optimal value of Z for the LP problem is not better than the current lower bound. The branch-and-bound algorithm continues to select a node for further branching until all the nodes have been fathomed. The fathomed node with the largest value of Z gives the optimal solution to the mixed-integer program. Hence, the efficiency of the branch-and-bound algorithm depends on how soon the successive nodes are fathomed. The fathoming conditions 1 and 2 generally take considerable time to reach. Condition 3 cannot be used until a lower bound for the MIP problem is found. However, a lower bound is not available until a feasible (integer) solution to the MIP problem is obtained (condition 1). Hence it is always helpful if a feasible integer solution to the MIP problem can be found before the start of the branch-and-bound procedure. This will provide the initial lower bound to the MIP problem until a better lower bound is found by the branch-and-bound algorithm. In many practical problems, the present way of operating the system may provide an initial solution. 11.1.3

Guidelines on Problem Formulation and Solution

The solution time for solving the integer programming problem is very sensitive to the way the problem is formulated initially. From practical experience in solving a number of integer programs by the branch-and-bound method, the IBM research staff has come up with some suggestions on model formulations. We describe them briefly here. The following guidelines should not be considered as restrictions but rather as suggestions that have often reduced computational time in practice. 1. Keep the number of integer variables as small as possible. One way to do this is to treat all integer variables whose values will be at least 20 as continuous variables. 2. Provide a good (tight) lower and upper bound on the integer variables when possible. 3. Unlike the general LP problem, the addition of new constraints to an MIP problem will generally reduce the computational time, especially when the new constraints contain integer variables.

11.1

INTEGER PROGRAMMING

493

4. If there is no critical need to obtain an exact optimal integer solution, then considerable savings in computational time may be obtained by accepting the first integer solution that is 3 percent of the continuous optimum. In other words, for a maximization problem we can terminate the branch-and-bound procedure whenever Upper bound ⫺ lower bound ⬍ 0.03 Upper bound 5. The order in which the integer variables are chosen for branching affects the solution time. It is recommended that the integer variables be processed in a priority order based on their economic significance and user experience. The branch-and-bound approach can also be used to solve nonlinear integer programming problems. Gupta and Ravindran [7] have implemented the branch-and-bound method with the generalized reduced gradient algorithm with some success. Solution of 0–1 Problems. For integer programming problems involving 0–

1 integer variables, there exist special-purpose algorithms for efficient solutions. These are generally based on an implicit enumeration method. For a good discussion of these algorithms, the reader is referred to Plane and McMillan [4] and Murty [5]. Software Packages. A number of software packages for solving MIP prob-

lems are available for mainframes, workstations, and PCs. The computational efficiency of the MIP software depends largely on the efficiency of the LP software available to solve LP subproblems at each node. Hence, progress in solving large MIP problems has gone hand in hand with advances in LP solution technology. Some notable commercial MIP software include IBM’s OSL (Optimization Subroutine Library), CPLEX mixed-integer optimizer, and LINDO. They are available for mainframes, workstations, and PCs. Solvers that can operate directly on spreadsheets are also available now. For example, Microsoft Excel contains a general-purpose optimizer for solving linear, integer and nonlinear programming problems. It has been developed by Frontline Systems, which also offers a more powerful Premium Solver for larger problems (see Albright [8]). Grossman [9] has a survey of other spreadsheet add-ins that are available. For a detailed description of all the available software for MIP problems and their capabilities, refer to the 1994 survey in OR/MS Today by Saltzman [10]. Another excellent source of reference is the resource handbook on linear and discrete optimization by Sharda [11]. For a detailed discussion of integer programming applications and algorithms, see references 5, 6, and 12.

494

11.2


QUADRATIC PROGRAMMING

A QP problem is an optimization problem involving a quadratic objective function and linear constraints. It can be stated in a general form as

冘 n

Minimize

ƒ(x) ⫽

冘冘x q x n

cj xj ⫹

n

ij

ij ij

i⫽1 j⫽1

j⫽1

冘a x ⫽b n

Subject to

ij j

i

for i ⫽ 1, . . . , m

j⫽1

xj ⭓ 0 In vector–matrix notation, it may be written as Minimize

ƒ(x) ⫽ cx ⫹ x⬘Qx

Subject to Ax ⫽ b x⭓0

(11.1) (11.2) (11.3)

where A is an m ⫻ n matrix of constraint coefficients, b is an m ⫻ 1 column vector, c is a 1 ⫻ n row vector, Q is an n ⫻ n matrix of quadratic form, and x is an n ⫻ 1 vector of design variables. 11.2.1

Applications of Quadratic Programming

We shall now illustrate some of the applications of QP models. Example 11.4

Portfolio Selection Problem

An important problem faced by financial analysts in investment companies (banks, mutual funds, insurance companies) is the determination of an optimal investment portfolio. A portfolio specifies the amount invested in different securities which may include bonds, common stocks, bank CDs, treasury notes, and others. Because of the economic significance of the problem and its complexity, a number of mathematical models have been proposed for analyzing the portfolio selection problem. Assume we have N securities for possible investment and are interested in determining the portion of available capital C that should be invested in each of the securities for the next investment period. Let the decision variables be denoted by xj, j ⫽ 1, 2, . . . , N, which represent the dollar amount of capital invested in security j. We then have the following system constraints on the decision variables:

11.2


495

x1 ⫹ x2 ⫹ 䡠䡠䡠 ⫹ xN ⭐ C xj ⭓ 0

j ⫽ 1, . . . , N

Suppose we have historical data on each security for the past T years, which give the price fluctuations and dividend payments. We can then estimate the return on investment from each security from past data. Let rj(t) denote the total return per dollar invested in security j during year t. Then rj(t) ⫽

pj(t ⫹ 1) ⫺ pj(t) ⫹ dj(t) pj(t)

where pj(t) is the price of security j at the beginning of year t and dj(t) is the total dividends received in year t. Note that the values of rj(t) are not constants and can fluctuate widely from year to year. In addition, rj(t) may be positive, negative, or zero. Hence, to assess the investment potential of security j, we can compute the average or expected return from security j per dollar invested, denoted by ␮j as ␮j ⫽

1 T

冘 r (t) T

j

t⫽1

The total expected return from the investment portfolio is then given by

冘␮x ⫽␮ x N

E⫽

T

j j

j⫽1

where ␮T ⫽ (␮1, . . . , ␮N) and x ⫽ (x1, . . . , xN)T. Model I. A simple optimization model for the investor’s problem is to max-

imize the total expected return subject to constraints on investment goals as follows:

冘␮x N

Maximize

Z⫽

j j

j⫽1

冘x ⭐C N

Subject to

j

xj ⭓ 0

j⫽1

A number of policy constraints may also be imposed on the portfolio. Most investment companies limit the amount that can be invested in common stocks, whose returns are subject to wide variations. This can be expressed as

496


冘 x ⭐b j

1

j僆J1

where J1 represents the common stock securities and b1 is the maximum investment ($) in common stocks. Many investment companies also require a certain amount in ready cash or fluid state to meet withdrawal requests from customers. This may be expressed as

冘 x ⭓b j

2

j僆J2

where J2 represents those securities in fluid state (e.g., savings accounts, checking accounts) and b2 is the minimum cash reserve required. Similarly, one can introduce a number of such policy constraints on the investment portfolio. A major drawback of this simple LP model is that it completely ignores the risk associated with the investment. Hence, it can lead to an investment portfolio that may have a very high average return but also has a very high risk associated with it. Because of the high risk, the actual return may be far less than the average value! Model II. In this model, we shall incorporate the ‘‘risk’’ associated with each

security. Some securities, such as ‘‘speculative stocks,’’ may have a larger price appreciation possibility and hence a greater average return, but they may also fluctuate more in value, increasing its risk. On the other hand, ‘‘safe’’ investments, such as savings accounts and bank CDs, may have smaller returns on investments. We can measure the investment risk of a security by the amount of fluctuation in total return from its average value during the past T years. Let ␴ 2jj denote the investment risk or variance of security j, which can be computed as ␴ 2jj ⫽

1 T

冘 [r (t) ⫺ ␮ ] T

2

j

j

t⫽1

In addition to the variance, a group of securities may be tied to a certain aspect of the economy, and a downturn in that economy would affect the prices of all those securities in that group. Automobile stocks, utility stocks, and oil company stocks are some examples of such security groups, which may rise or fall together. To avoid this kind of risk, we must diversify the investment portfolio so that funds are invested in several different security groups. This can be done by computing the relationships in rate of return between every pair of securities. Such a relationship, known as covariance, is denoted by ␴ 2ij and can be computed from past data:

11.2

␴ 2ij ⫽

1 T


497

冘 [r (t) ⫺ ␮ ][r (t) ⫺ ␮ ] T

i

i

j

j

t⫽1

Note that when i ⫽ j, the above equation reduces to the variance of security j. Thus, we can measure the variance or investment risk of a portfolio by

冘冘 ␴ x x ⫽ x Qx N

V⫽

N

2 ij i j

T

i⫽1 j⫽1

where Q(N⫻N) ⫽ [␴ 2ij] is the variance–covariance matrix of the N securities. A rational investor may be interested in obtaining a certain average return from the portfolio at a minimum risk. This can be formulated as Minimize

Z ⫽ xTQx

Subject to

冘x ⭐C N

j

j⫽1

xj ⭓ 0 ␮Tx ⭓ R

where R is the minimum average return desired from the portfolio. Additional policy constraints of the type discussed in model I could be added to the above model. In this risk-averse model, the constraints of the problem are linear but the objective is a quadratic function. Example 11.5

Constrained Regression Problem

The classical regression problem involves determining the coefficients of a regression equation that gives a relationship between a dependent variable and a set of independent variables. The basic least-squares regression problem may be stated as Minimize

eTIe

Subject to Y ⫽ X␤ ⫹ e where ␤ is an m ⫻ 1 vector of regression coefficients to be estimated, e is an n ⫻ 1 vector of error variables, Y is an n ⫻ 1 vector of observations on the dependent variable, and X is an n ⫻ m matrix of observations on the independent variables. It is clear that the classical regression problem is a quadratic program. In many practical problems a number of additional constraints may be imposed based on prior information about the problem. These may involve

498


restrictions on the regression coefficients ␤j (j ⫽ 1, . . . , m) such as ␤j ⭓ 0 and 兺 ␤j ⫽ 1. One may also decide to weight the error terms ei’s unevenly. In other words, the objective function may be written as 兺i wi e2i . Thus a QP approach to regression problems allows more flexibility in the model. For more applications of QP, the reader is referred to McCarl, Moskowitz, and Furtan [13]. 11.2.2

Kuhn–Tucker Conditions

Comparing the quadratic program [Eqs. (11.1)–(11.3)] to the general nonlinear program given in Chapter 5 [Section 5.5, Eqs. (5.10)–(5.12)], we get ƒ(x) ⫽ cx ⫹ xTQx g(x) ⫽ x ⭓ 0 h(x) ⫽ Ax ⫺ b ⫽ 0 Using Eqs. (5.13)–(5.17), the associated Kuhn–Tucker conditions to the quadratic program can be written as c ⫹ xT(Q ⫹ QT) ⫺ u ⫺ vA ⫽ 0 Ax ⫽ b

(11.4)

x⭓0

(11.5)

ux ⫽ 0 u⭓0

(11.6)

v unrestricted in sign

(11.7)

We shall illustrate the Kuhn–Tucker conditions with an example. Example 11.6

Minimize

ƒ(x) ⫽ ⫺6x1 ⫹ 2x 21 ⫺ 2x1x2 ⫹ 2x 22

Subject to x1 ⫹ x2 ⫽ 2 x1, x2 ⭓ 0 For this problem, c ⫽ (⫺6, 0)

Q⫽

冉

冊

2 ⫺1 2

⫺1

The Kuhn–Tucker conditions are given by

A ⫽ (1, 1)

b ⫽ (2)

11.3

冉

(⫺6, 0) ⫹ (x1, x2)

4

⫺2

COMPLEMENTARY PIVOT PROBLEMS

冊

⫺2

4

499

⫺ (u1, u2) ⫺ v1(1, 1) ⫽ 0

or ⫺6 ⫹ 4x1 ⫺ 2x2 ⫺ u1 ⫺ v1 ⫽ 0

0 ⫺ 2x1 ⫹ 4x2 ⫺ u2 ⫺ v1 ⫽ 0 x1 ⫹ x2 ⫽ 2 u1x1 ⫽ 0 u1, u2 ⭓ 0

x1, x2 ⭓ 0 u2 x2 ⫽ 0

v1 unrestricted in sign

Since the constraints of the quadratic program are linear, the constraint qualification is always satisfied and the Kuhn–Tucker necessity theorem applies. (See Theorem 5.2 of Section 5.6.) Hence, the Kuhn–Tucker conditions are necessary for the optimality of the quadratic program. Similarly, when the matrix of quadratic form Q is positive definite or positive semidefinite, the objective function is a convex function and the Kuhn– Tucker conditions become sufficient for the optimality of the quadratic program. Hence, when Q is positive definite or positive semidefinite, it is sufficient to solve the Kuhn–Tucker conditions to find an optimal solution to the quadratic program. Note that Eqs. (11.4) and (11.5) represent a system of simultaneous equations that can be solved as a linear program. Hence, one of the earliest procedures for solving a quadratic program developed by Wolfe [14] solves Eqs. (11.4), (11.5), and (11.7) as a linear program using the phase I simplex method by creating artificial variables if necessary (see Chapter 4). Equation (11.6) is implicitly handled in the simplex algorithm using a ‘‘restricted basis entry rule’’ (see Chapter 8, Section 8.2). This would prohibit a nonbasic variable uj from entering the basis if the corresponding xj is in the basis at a positive value. Unfortunately, Wolfe’s method fails to converge when the matrix Q is positive semidefinite (instead of positive definite). A more efficient and simple method of solving the Kuhn–Tucker conditions (11.4)–(11.7) is the complementary pivot method developed by Lemke [15]. In an experimental study, Ravindran and Lee [16] have shown that this method is computationally more attractive than most of the methods available for solving QP problems when Q is positive semidefinite. The complementary pivot method will be discussed in the next section as a general procedure for solving a special class of problems known as complementary problems which includes linear and quadratic programs. 11.3


Consider the general problem of finding a nonnegative solution to a system of equations of the following form: Find vectors w and z such that

500


w ⫽ Mz ⫹ q w⭓0

(11.8)

z⭓0

w⬘z ⫽ 0

(11.9) (11.10)

where M is an n ⫻ n square matrix and w, z, q are n-dimensional column vectors. The above problem is known as a complementary problem. Note that there is no objective function to minimize or maximize in this formulation. Condition (11.8) represents a system of simultaneous linear equations; condition (11.9) requires that the solution to condition (11.8) be nonnegative; condition (11.10) implies wi zi ⫽ 0 for all i ⫽ 1, 2, . . . , n since wi, zi ⭓ 0. Thus we have a single nonlinear constraint. Example 11.7

Consider a problem with

冢

1 2 3 M⫽ 4 5 6 6 7 8

冣

and

冢冣

2 q ⫽ ⫺5 ⫺3

The complementary problem is given as follows: Find w1, w2, w3, z1, z2, z3 such that w1 ⫽ z1 ⫹ 2z2 ⫹ 3z3 ⫹ 2 w2 ⫽ 4z1 ⫹ 5z2 ⫹ 6z3 ⫺ 5 w3 ⫽ 6z1 ⫹ 7z2 ⫹ 8z3 ⫺ 3 w1, w2, w3, z1, z2, z3 ⭓ 0 w1z1 ⫹ w2z2 ⫹ w3z3 ⫽ 0 Consider a convex QP problem of the form

11.3


501

ƒ(x) ⫽ cx ⫹ xTQx

Minimize

Subject to Ax ⭓ b x⭓0 where Q is an n ⫻ n matrix. Assume Q is symmetric and is positive definite or positive semidefinite. In matrix notation, the Kuhn–Tucker optimality conditions to the above convex quadratic program can be written as follows: Find vectors x, y, u, s such that u ⫽ 2Qx ⫺ ATy ⫹ cT s ⫽ Ax ⫺ b x, y, u, s ⭓ 0

uTx ⫹ sTy ⫽ 0

Comparing the above system of equations to the complementary problem, we note that w⫽

冉冊 u s

z⫽

冉冊 x y

M⫽

冉

冊

2Q ⫺AT A 0

q⫽

冉冊 cT ⫺b

Thus an optimal solution to the convex quadratic program may be obtained by solving the equivalent complementary problem shown above. It should be again noted that the matrix M is positive semidefinite since Q is positive definite or positive semidefinite. The reduction of a linear program as a complementary problem can be easily achieved by setting the matrix Q to zero. Hence, complementary pivot algorithms can be used to solve linear programs as well. Example 11.8

Consider a convex QP problem: Minimize

ƒ(x) ⫽ ⫺6x1 ⫹ 2x 21 ⫺ 2x1x2 ⫹ 2x 22

Subject to ⫺x1 ⫺ x2 ⭓ ⫺2 x1, x2 ⭓ 0 For this problem,

502


A ⫽ (⫺1, ⫺1)

b ⫽ (⫺2)

cT ⫽

冉冊 ⫺6

0

and Q⫽

冋

2 ⫺1 2

⫺1

册

The equivalent complementary problem is given by M(3⫻3) ⫽

冉

冊冢

4 Q ⫹ QT ⫺AT ⫽ ⫺2 A 0 ⫺1

⫺2

冣

1 4 1 ⫺1 0

and q⫽

冉冊冢冣 cT ⫽ ⫺b

⫺6

0 2

The values of z1 and z2 correspond to the optimal values of x1 and x2 since the matrix of the quadratic form (Q) is positive definite. Complementary Pivot Method. Consider the complementary problem: Find

vectors w and z such that w ⫽ Mz ⫹ q w, z ⭓ 0

w⬘z ⫽ 0

Definitions A nonnegative solution (w, z) to the system of equation w ⫽ Mz ⫹ q is called a feasible solution to the complementary problem. A feasible solution (w, z) to the complementary problem that also satisfies the complementarity condition wTz ⫽ 0 is called a complementary solution. The condition wTz ⫽ 0 is equivalent to wi zi ⫽ 0 for all i. The variables wi and zi for each i is called a complementary pair of variables. Note that if the elements of the vector q are nonnegative, then there exists an obvious complementary solution given by w ⫽ q, z ⫽ 0. Hence the complementary problem is nontrivial only when at least one of the elements of q is negative. This means that the initial basic solution given by w ⫽ q, z ⫽ 0 is infeasible to

11.3


503

the complementary problem even though it satisfies the complementary condition wTz ⫽ 0. The complementary pivot method (Lemke [15]) starts with the infeasible basic solution given by w ⫽ q, z ⫽ 0. To make the solution nonnegative, an artificial variable z0 is added at a sufficiently positive value to each of the equations in the w ⫺ Mz ⫽ q system, so that the right-hand-side constants (qi ⫹ z0) become nonnegative. The value of z0 will be the absolute value of the most negative qi. We now have a basic solution given by wi ⫽ qi ⫹ z0

zi ⫽ 0

for all i ⫽ 1, . . . , n

and z0 ⫽ ⫺min(qi) Note that, as previously discussed, even though this solution is nonnegative, satisfies the constraints, and is complementary (wi zi ⫽ 0), it is not feasible because of the presence of the artificial variable z0 at a positive value. We shall call such a solution an almost complementary solution. The first step in the complementary pivot method is to find an almost complementary solution by augmenting the original system of equations (w ⫽ Mz ⫹ q) by an artificial variable z0 as follows: w ⫺ Mz ⫺ ez0 ⫽ q w, z, z0 ⭓ 0 wTz ⫽ 0 where e(n⫻1) ⫽ (1, 1, . . . , 1)⬘ Thus, the initial tableau becomes Basis

w1

w1 ws wn

1

䡠䡠䡠

ws

䡠䡠䡠

wn

z1

1

⫺m11 ⫺ms1 ⫺mn1

1

䡠䡠䡠

zs ⫺m1s

mss ⫺mns

䡠䡠䡠

zn

z0

q

⫺m1n ⫺msn ⫺mnn

⫺1 ⫺1 ⫺1

q1 qs qn

where the mij’s are the elements of the M matrix. Step 1. To determine the initial almost complementary solution, the variable z0 is brought into the basis, replacing the basis variable with the most

504


negative value. (Let qs ⫽ min qi ⬍ 0.) This implies that z0 replaces ws from the basis. Performing the necessary pivot operation yields Tableau 1. where qs⬘ ⫽ ⫺qs msj⬘ ⫽

⫺msj

qi⬘ ⫽ qi ⫺ qs

for all i ⫽ s

⫽ msj

for all j ⫽ 1, . . . , n

m⬘ij ⫽ ⫺mij ⫹ msj

for all j ⫽ 1, . . . , n

⫺1

(11.11) (11.12) i⫽s

and

(11.13)

Note: 1. q⬘i ⭓ 0, i ⫽ 1, . . . , n. 2. The basic solution w1 ⫽ q1⬘, . . . , ws⫺1 ⫽ qs⫺1 ⬘ , z0 ⫽ q⬘s , ws⫹1 ⫽ q⬘s⫹1, . . . , wn ⫽ qn⬘, and all other variables initially zero is an almost complementary solution. 3. The almost complementary solution becomes a complementary solution as soon as the value of z0 is reduced to zero. In essence, the complementary pivot algorithm proceeds to find a sequence of almost complementary solutions (tableaus) until z0 becomes zero. To do this, the basis changes must be done in such a way that the following conditions are met: (a) The complementarity between the variables must be maintained (i.e., wi zi ⫽ 0 for all i ⫽ 1, . . . , n). (b) The basic solution remains nonnegative (i.e., the right-hand-side constants must be nonnegative in all tableaus). Step 2. To satisfy condition (a), we observe that the variables ws and zs are both out of the basis in Tableau 1. As long as either one of them is made basic, the complementarity between w and z variables will still be maintained. Since ws just came out of the basis, the choice is naturally to bring zs into the basis. Thus we have a simple rule for selecting the nonbasic variable to enter the basis in the next tableau. It is always the complement

Tableau 1

Basis

w1

w1 z0 wn

1 0 0

䡠䡠䡠

ws ⫺1 ⫺1 ⫺1

䡠䡠䡠

wn

z1

0 0 1

m⬘11 m⬘s1 m⬘n1

䡠䡠䡠

zs m⬘1s m⬘ss m⬘ns

䡠䡠䡠

zn

z0

q

m⬘1n m⬘sn m⬘nn

0 1 0

q⬘1 q⬘s q⬘n

11.3


505

of the basic variable that just left the basis in the last tableau. This is called the complementary rule. After selecting the variable to enter the basis, we have to determine the basic variable to leave. This is done by applying the minimum-ratio test similar to the one used in the LP simplex method so that condition (b) is satisfied. Therefore, to determine the variable to leave the basis, the following ratios are formed: q⬘i m⬘is

for those i ⫽ 1, . . . , n for which mis⬘ ⬎ 0

Let

冉冊

q⬘i q⬘k ⫽ min mks ⬘ m⬘is mis⬎0

(11.14)

This implies that the basic variable wk leaves the basis to be replaced by zs. We now obtain the new tableau by performing the pivot operation with m⬘ks as the pivot element. Step 3. Since wk left the basis, the variable zk is brought into the basis by the complementary rule, and the basis changes are continued as before. Termination of the Algorithm 1. The minimum ratio is obtained in row s, and z0 leaves the basis. The resulting basic solution after performing the pivot operation is the complementary solution. 2. The minimum-ratio test fails, since all the coefficients in the pivot column are nonpositive. This implies that no solution to the complementary problem exists. In this case, we say that the complementary problem has a ray solution. Remarks 1. It has been shown that the complementary pivot method always terminates with a complementary solution in a finite number of steps whenever (a) all the elements of M are positive or (b) M has positive principal determinants (includes the case where M is positive definite). 2. The most important application of complementary pivot theory is in solving convex QP problems. We have seen that under these cases M is a positive-semidefinite matrix. It has been proved that whenever M is positive semidefinite, the algorithm will terminate with a complementary solution if one exists for that problem. In other words, termi-

506


nation 2 implies that the given linear program or quadratic program has no optimal solution. For a proof of the above remarks, the reader is referred to Cottle and Dantzig [17]. Let us illustrate the complementary pivot method for solving a convex quadratic program using Example 11.8. Minimize

ƒ(x) ⫽ ⫺6x1 ⫹ 2x 12 ⫺ 2x1x2 ⫹ 2x 22

Subject to ⫺x1 ⫺ x2 ⭓ ⫺2 x1, x2 ⭓ 0 The equivalent complementary problem is

冢

4 ⫺2 1 4 1 M ⫽ ⫺2 (3⫻3) ⫺1 ⫺1 0

冣

and

q⫽

冢冣 ⫺6

0 2

Since all the elements of q are not nonnegative, an artificial variable z0 is added to every equation. The initial tableau is given as Tableau 2. The initial basic solution is w1 ⫽ ⫺6, w2 ⫽ 0, w3 ⫽ 2, z1 ⫽ z2 ⫽ z3 ⫽ z0 ⫽ 0. An almost complementary solution is obtained by replacing w1 by z0 and using Eqs. (11.11)–(11.13) as shown in Tableau 3. The almost complementary solution is given by z0 ⫽ 6, w2 ⫽ 6, w3 ⫽ 8, z1 ⫽ z2 ⫽ z3 ⫽ w1 ⫽ 0. Since the complementary pair (w1, z1) is out of the basis, either w1 or z1 can be made a basic variable without affecting the complementarity between all pairs of variables (wi zi ⫽ 0). Since w1 just left the basis, we bring z1 into the basis. Applying the minimum-ratio test [Eq. (11.14)], we obtain the ratios as –64 , –66 , –85 ). This implies that z1 replaces w2 in the basis. Tableau 4 gives the new almost complementary solution after the pivot operation. By applying the complementary rule, z2 is selected as the next basic variable (w2 just left the basis). The minimum-ratio test determines w3 as the basic variable to leave. The next almost complementary solution after the pivot operation is shown in Tableau 5. By the complementary rule, z3 becomes the next basic variable. Application of the minimum-ratio test results in the replacement of z0 from the basis. This Tableau 2

Basis

w1

w2

w3

z1

z2

z3

w1 w2 w3

1 0 0

0 1 0

0 0 1

⫺4

2 ⫺4 1

⫺1 ⫺1

2 1

0

z0

q

⫺1 嘷

⫺6

⫺1 ⫺1

0 2

11.4

507

GOAL PROGRAMMING

Tableau 3

Basis

w1

w2

w3

z1

z2

z3

z0

q

z0 w2 w3

⫺1 ⫺1 ⫺1

0 1 0

0 0 1

嘷 6

4

⫺2 ⫺6 ⫺1

1 0 1

1 0 0

6 6 8

5

implies that the next tableau will correspond to a complementary solution as shown in Tableau 6. The complementary solution is given by z1 ⫽ –32 , z2 ⫽ –12 , z3 ⫽ 1, w1 ⫽ w2 ⫽ w3 ⫽ 0. Hence the optimal solution to the given quadratic program becomes x*1 ⫽ –23

1 x* 2 ⫽ – 2

11 ƒ(x*) ⫽ ⫺–– 2

A computer program implementing the complementary pivot method is given by Ravindran [18]. Experimental studies with this code by Ravindran [19, 20] and Ravindran and Lee [16] have shown the superiority of the complementary pivot method for solving linear programs and convex quadratic programs, respectively. In the QP study [16], the convex simplex method (see Section 9.2.1) was a close second to the complementary pivot method for solving the convex quadratic programs. For the nonconvex quadratic programs, the convex simplex method should be preferred, since it guarantees convergence to a local optimum. On the other hand, for nonconvex problems the complementary pivot method may terminate with a ray solution (see termination 2) and fail to identify even a local optimum. 11.4

GOAL PROGRAMMING [3]

All our discussion of optimization methods until now has considered only one criterion or objective as the performance measure to define the optimum. It is not possible to find a solution that, say, simultaneously minimizes cost, maximizes reliability, and minimizes energy utilization. This again is an important simplification of reality because in many practical situations it would be desirable to achieve a solution that is ‘‘best’’ with respect to a number of different criteria. One way of treating multiple competing criteria is to select one criterion as primary and the other criteria as secondary. The primary Tableau 4

Basis

w1

w2

w3

z1

z2

z3

z0

q

z0 z1 w3

⫺–13 ⫺–16 ⫺–16

⫺–23 –16 ⫺–56

0 0 1

0 1 0

2 ⫺1

1 0 1

1 0 0

2 1 3

嘷 4

508


Tableau 5

Basis z0 z1 z2

w1

w2

w3

z1

z2

⫺–14 5 ⫺–– 24 1 ⫺–– 24

⫺–14 1 ⫺–– 24 5 ⫺–– 24

⫺–12 –14 –14

0 1 0

0 0 1

z3 – 嘷 1 2 1 4 1 4

– –

z0

q

1 0 0

–12 –74 –34

criterion is then used as the optimization objective function, while the secondary criteria are assigned acceptable minimum and maximum values and are treated as problem constraints. However, if careful considerations were not given while selecting the acceptable levels, a feasible design that satisfies all the constraints may not exist. This problem is overcome by a technique called goal programming, which is fast becoming a practical method for handling multiple criteria. In goal programming, all the objectives are assigned target levels for achievement and a relative priority on achieving these levels. Goal programming treats these targets as goals to aspire for and not as absolute constraints. It then attempts to find an optimal solution that comes as ‘‘close as possible’’ to the targets in the order of specified priorities. In this section, we shall discuss how to formulate goal programming models, their solution methods, and their applications. Before we discuss the formulation of goal programming problems, we should discuss the difference between the terms real constraints and goal constraints (or simply goals) as used in goal programming models. The real constraints are absolute restrictions on the decision variables, while the goals are conditions one would like to achieve but are not mandatory. For instance, a real constraint given by x1 ⫹ x2 ⫽ 3 requires all possible values of x1 ⫹ x2 to always equal 3. As opposed to this, a goal requiring x1 ⫹ x2 ⫽ 3 is not mandatory, and we can choose values of x1 ⫹ x2 ⭓ 3 as well as x1 ⫹ x2 ⭐ 3. In a goal constraint, positive and negative deviational variables are introduced as follows:

Tableau 6

Basis z3 z1 z2

w1

w2

w3

z1

z2

z3

z0

q

⫺–12 1 ⫺–– 12 1 ⫺–– 12

⫺–12 1 ⫺–– 12 1 ⫺–– 12

⫺1 –21 –21

0 1 0

0 0 1

1 0 0

2

1

⫺–12 ⫺–12

–32 –12

11.4

x1 ⫹ x2 ⫹ d1⫺ ⫺ d1⫹ ⫽ 3

GOAL PROGRAMMING

509

d1⫹, d1⫺ ⭓ 0

⫹ Note that, if d⫺ 1 ⬎ 0, then x1 ⫹ x2 ⬍ 3, and if d1 ⬎ 0, then x1 ⫹ x2 ⬎ 3. By ⫺ ⫹ ⫺ assigning suitable weights w1 and w1 on d1 and d1⫹ in the objective function, the model will try to achieve the sum x1 ⫹ x2 as close as possible to 3. If the goal were to satisfy x1 ⫹ x2 ⭓ 3, then only d⫺ 1 is assigned a positive weight in the objective, while the weight on d⫹ is set to zero. 1 The general goal programming model can be expressed as follows.

冘 (w d m

Minimize

Z⫽

⫹ ⫹ i i

⫺ ⫹ w⫺ i di )

(11.15)

i⫽1

冘a x ⫹d n

Subject to

⫺ i

ij j

j⫽1

⫺ di⫹ ⫽ bi

⫹ xj, d⫺ i , di ⭌ 0

for all i

(11.16)

for all i and j

(11.17)

Equation (11.15) represents the objective function, which minimizes the weighted sum of the deviational variables. The system of equations (11.16) represents the goal constraints, relating the decision variables (xj) to the targets (bi), and the set of inequalities (11.17) represents the standard nonnegativity restrictions on all variables. ⫺ If the relative weights (w⫹ i and wi ) can be specified by the management, then the goal programming problem reduces to a simple linear program. Unfortunately, it is difficult or almost impossible in many cases to secure a numerical approximation to the weights. In reality, goals are usually incompatible (i.e., incommensurable) and some goals can be achieved only at the expense of some other goals. Hence, goal programming uses ordinal ranking or preemptive priorities to the goals by assigning incommensurable goals to different priority levels and weights to goals at the same priority level. The objective function (11.15) takes the form Minimize Z ⫽

冘 P 冘 (w d

⫹ ⫹ ik i

k

k

⫺ ⫹ w⫺ ik di )

i

where Pk represents priority k with the assumption that Pk is much larger than ⫺ Pk⫹1 and w⫹ ik and wik are the weights assigned to the ith deviational variables at priority k. In this manner, lower priority goals are considered only after attaining the higher priority goals. Thus, preemptive goal programming or simply goal programming is essentially a sequential optimization process, in which successive optimizations are carried out on the alternate optimal solutions of the previously optimized goals at higher priority. Goal programming formulation is illustrated by the next two examples.

510


Example 11.9

Suppose a company has two machines for manufacturing a product. Machine 1 makes 2 units per hour, while machine 2 makes 3 per hour. The company has an order for 80 units. Energy restrictions dictate that only one machine can operate at one time. The company has 40 hr of regular machining time, but overtime is available. It costs $4.00 to run machine 1 for 1 hr, while machine 2 costs $5.00 per hour. The company’s goals, in order of importance are as follows: 1. 2. 3. 4.

Meet the demand of 80 units exactly Limit machine overtime to 10 hr. Use the 40 hr of normal machining time. Minimize costs.

Formulation. Letting xj represent the number of hours machine j is operating, the goal programming model is Minimize

⫹ ⫹ ⫺ ⫹ ⫹ Z ⫽ P1(d⫺ 1 ⫹ d1 ) ⫹ P2d3 ⫹ P3(d2 ⫹ d2 ) ⫹ P4d4

Subject to

⫹ 2x1 ⫹ 3x2 ⫹ d⫺ 1 ⫺ d1 ⫽ 80

(11.18)

⫹ x1 ⫹ x2 ⫹ d⫺ 2 ⫺ d2 ⫽ 40

(11.19)

d⫹ 2 ⫹

(11.20)

⫹ d3⫺ ⫺ d⫹ 3 ⫽ 10

⫹ 4x1 ⫹ 5x2 ⫹ d⫺ 4 ⫺ d4 ⫽ 0 ⫹ xi, d⫺ i , di ⭌ 0

(11.21)

for all i

where P1, P2, P3, and P4 represent the preemptive priority factors such that P1 ⬎ ⬎ P2 ⬎ ⬎ P3 ⬎ ⬎ P4. Note that the target for cost is set at an unrealistic level of zero. Since the goal is to minimize d⫹ 4 , This is equivalent to minimizing the total cost of production. In this formulation, Eq. (11.20) does not conform to the general model given by Eqs. (11.15)–(11.17), where no goal constraint involves a deviational ⫺ variable defined earlier. However, if d⫹ 2 ⬎ 0, then d2 ⫽ 0, and from Eq. (11.19), we get d⫹ 2 ⫽ x1 ⫹ x2 ⫺ 40 which when substituted into Eq. (11.20) yields ⫹ x1 ⫹ x2 ⫹ d⫺ 3 ⫺ d3 ⫽ 50

(11.22)

11.4

GOAL PROGRAMMING

511

Thus, Eq. (11.22) can replace Eq. (11.20) and the problem fits the general model, where each deviational variable appears in only one goal constraint and has at most one positive weight in the objective function. Example 11.10 Economic Machining Problem with Two Competing Objectives [21]

Consider a single-point, single-pass turning operation in metal cutting wherein an optimum set of cutting speed and feed rate is to be chosen which balances the conflict between metal removal rate and tool life as well as being within the restrictions of horsepower, surface finish, and other cutting conditions. In developing the mathematical model of this problem, the following constraints will be considered for the machining parameters: Constraint 1: Maximum Permissible Feed ƒ ⬉ ƒmax where ƒ is the feed in inches per revolution. A cutting force restriction or surface finish requirements usually determine ƒmax. [See Armarego and Brown [22].] Constraint 2: Maximum Cutting Speed Possible. If ␯ is the cutting speed in surface feet per minute, then ␯ ⬉ ␯max

1 ␯max ⫽ –– 12 ␲DNmax

where D ⫽ mean work piece diameter, in. Nmax ⫽ maximum spindle speed available on machine, rpm Constraint 3: Maximum Horsepower Available. If Pmax is the maximum horsepower available at the spindle, then ␯ƒ␣ ⭐

Pmax(33,000) ct d␤c

where ␣, ␤, and ct are constants [22]. The depth of cut in inches, dc, is fixed at a given value. For a given Pmax, ct, ␤, and dc, the right-hand side of the above constraint will be a constant. Hence, the horsepower constraint can be simply written as ␯ƒ␣ ⬉ const

512


Constraint 4: Nonnegativity Restrictions on Feed Rate and Speed ␯, ƒ ⭌ 0

In optimizing metal cutting there are a number of optimality criteria which can be used. Suppose we consider the following objectives in our optimization: (i) Maximize metal removal rate (MRR) and (ii) maximize tool life (TL). The expression MRR is MRR ⫽ 12␯ƒdc in.3 /min The TL for a given depth of cut is given by TL ⫽

A ␯ ƒ

1 / n 1 / n1

where A, n, and n1 are constants. We note that the MRR objective is directly proportional to feed and speed, while the TL objective is inversely proportional to feed and speed. In general, there is no single solution to a problem formulated in this way, since MRR and TL are competing objectives and their respective maxima must include some compromise between the maximum of MRR and the maximum of TL. Goal Programming Model [23]. Management considers that a given single-

point, single-pass turning operation will be operating at an acceptable efficiency level if the following goals are met as closely as possible: 1. The metal removal rate must be greater than or equal to a given rate M1 (in.3 /min). 2. The tool life must equal T1 (min). In addition, management requires that a higher priority be given to achieving the first goal than the second. The goal programming approach may be illustrated by expressing each of the goals as goal constraints as shown below. Taking the MRR goal first, ⫹ 12␯ƒdc ⫹ d⫺ 1 ⫺ d1 ⫽ M1

where d⫺ 1 represents the amount by which the MRR goal is underachieved and d⫹ 1 represents any overachievement of the MRR goal. Similarly, the TL goal can be expressed as

11.4

GOAL PROGRAMMING

513

A ⫹ ⫹ d⫺ 2 ⫺ d2 ⫽ T1 ␯1 / nƒ1 / n1 Since the objective is to have a MRR of at least M1, the objective function must be set up so that a high penalty will be assigned to the underachievement ⫹ variable d⫺ 1 . No penalty will be assigned to d1 . To achieve a TL of T1, pen⫺ ⫹ alties must be associated with both d2 and d2 so that both of these variables are minimized to their fullest extent. The relative magnitudes of these penalties must reflect the fact that the first goal is considered to be more important than the second. Accordingly, the goal programming objective function for this problem is ⫺ ⫹ Minimize Z ⫽ P1d⫺ 1 ⫹ P2(d2 ⫹ d2 )

where P1 and P2 are nonnumerical preemptive priority factors such that P1 ⬎ ⬎ P2 (i.e., P1 is infinitely larger than P2). With this objective function every effort will be made to completely satisfy the first goal before any attempt is made to satisfy the second. To express the problem as a linear goal programming problem, M1 is replaced by M2, where M2 ⫽

M1 12dc

The goal T1 is replaced by T2, where T2 ⫽

A T1

and logarithms are taken of the goals and constraints. The problem can then be stated as follows: ⫺ ⫹ Z ⫽ P1d⫺ 1 ⫹ P2(d2 ⫹ d2 )

Minimize Subject to (MRR) goal

⫹ log ␯ ⫹ log ƒ ⫹ d⫺ 1 ⫺ d1 ⫽ log M2

(TL goal)

⫹ 1/n log ␯ ⫹ 1/n1 log ƒ ⫹ d⫺ 2 ⫺ d2 ⫽ log T2

(ƒmax constraint)

log ƒ ⬉ log ƒmax

(Vmax constraint)

log ␯ ⬉ log ␯max

(HP constraint)

log ␯ ⫹ ␣ log ƒ ⬉ log const

⫹ ⫺ ⫹ log ␯, log ƒ, d⫺ 1 , d1 , d2 , d2 , ⭌ 0

514


We would like to emphasize again that the last three inequalities are real constraints on feed, speed, and HP that must be satisfied at all times, whereas the equations for MRR and TL are simply goal constraints. For a further discussion of this problem and its solution, see reference 23. Partitioning Algorithm. Goal programming problems can be solved efficiently by the partitioning algorithm developed by Arthur and Ravindran [24, 25]. It is based on the fact that the definition of preemptive priorities implies that higher order goals must be optimized before lower order goals are even considered. Their procedure consists of solving a series of linear programming subproblems by using the solution of the higher priority problem as the starting solution for the lower priority problem. The partitioning algorithm begins by solving the smallest subproblem S1, which is composed of those goal constraints assigned to the highest priority P1 and the corresponding terms in the objective function. The optimal tableau for this subproblem is then examined for alternate optimal solutions. If none exist, then the present solution is optimal for the original problem with respect to all of the priorities. The algorithm then substitutes the values of the decision variables into the goal constraints of the lower priorities to calculate their attainment levels, and the problem is solved. However, if alternate optimal solutions do exist, the next set of goal constraints (those assigned to the second highest priority) and their objective function terms are added to the problem. This brings the algorithm to the next largest subproblem in the series, and the optimization resumes. The algorithm continues in this manner until no alternate optimum exists for one of the subproblems or until all priorities have been included in the optimization. The linear dependence between each pair of deviational variables simplifies the operation of adding the new goal constraints to the optimal tableau of the previous subproblem without the need for a dual-simplex iteration. When the optimal solution to the subproblem Sk⫺1 is obtained, a variable elimination step is performed prior to the addition of goal constraints of priority k. The elimination step involves deleting all nonbasic columns which have a positive relative cost Cj ⬎ 0) in the optimal tableau of Sk⫺1 from further consideration. This is based on the well-known LP result that a nonbasic variable with a positive relative cost in an optimal tableau cannot enter the basis to form an alternate optimal solution. Figure 11.7 gives a flowchart of the partitioning algorithm. We now illustrate the partitioning algorithm using Example 11.9. The subproblem S1 for priority P1 to be solved initially is given below:

S1:

Minimize

Z1 ⫽ d1⫺ ⫹ d⫹ 1

Subject to

2x1 ⫹ 3x2 ⫹ d1⫺ ⫺ d1⫹ ⫽ 80 ⫹ x1, x2, d⫺ 1 , d1 ⭌ 0

11.4

GOAL PROGRAMMING

515

Figure 11.7. Flowchart of the partitioning algorithm.

The solution to subproblem S1 by the simple method is given in Table 11.1. However, alternate optima exist to subproblem S1 (the nonbasic variable x1 ⫺ has a relative cost of zero). Since the relative costs for d⫹ 1 and d1 are positive, they cannot enter the basis later; else they destroy the optimality achieved for priority 1. Hence, they are eliminated from the tableau from further consideration. We now add the goal constraint assigned to the second priority [Eq. (11.22)]: ⫹ x1 ⫹ x2 ⫹ d⫺ 3 ⫺ d3 ⫽ 50

Since x2 is a basic variable in the present optimal tableau (Table 11.1), we perform a row operation on the above equation to eliminate x2, and we get ⫹ 70 –13 x1 ⫹ d⫺ 3 ⫺ d3 ⫽ –– 3

(11.23)

516


Table 11.1

Solution to Subproblem S1

P1: cj

0

0

1

1

x1

x2

d⫺ 1

d⫹ 1

b

d

2

嘷 3

1

⫺1

80

P1: c Row

⫺2

⫺3 ↑

0

2

–23

1

–13

⫺–13

0

0

1

1

cB

Basis ⫹ 1

1

0

x2 P1: c Row

Z1 ⫽ 80 80 –– 3

Z1 ⫽ 0

Equation (11.23) is now added to the present optimal tableau after deleting ⫹ the columns corresponding to d⫺ 1 and d1 . This is shown in Table 11.2. The objective function of subproblem S2 is given by Minimize Z2 ⫽ d⫹ 3 Since the right-hand side of the new goal constraint [Eq. (11.22)] remained nonnegative after the row reduction [Eq. (11.23)], d⫺ 3 was entered as the basic variable in the new tableau (Table 11.2). If, on the other hand, the right-hand side had become negative, the row would be multiplied by ⫺1 and d⫹ 3 would become the new basic variable. Table 11.2 indicates that we have found an optimal solution to S2. Since alternate optimal solutions exist, we add the goal constraint and objective corresponding to priority 3. We also eliminate the column corresponding to d⫹ 3 . The goal constraint assigned to P3, given by ⫹ x1 ⫹ x2 ⫹ d⫺ 2 ⫺ d2 ⫽ 40

is added after eliminating x2. This is shown in Table 11.3. Now, x1 can enter the basis to improve the priority 3 goal, while maintaining the levels achieved for priorities 1 and 2. Then d⫺ 2 is replaced by x1 and the next solution becomes optimal for subproblem S3 (see Table 11.3). Moreover, the solution obtained is unique. Hence, it is not possible to improve the goal corresponding to priority 4, and we terminate the partitioning algorithm. It is only necessary

Table 11.2


P2: cj

0

0

0

1

cB

Basis

x1

x2

d

d⫹ 3

b

0 0

x2 d⫺ 3

–23 –13

1 0

0 1

0 ⫺1

80 –– 3 70 –– 3

0

0

0

1

P2: c Row

⫺ 3

Z2 ⫽ 0

11.4

Table 11.3

0 0 1

0

0

0

1

1

x1

x2

d⫺ 3

d⫺ 2

d⫹ 2

b

–23 –13 –13

1 0 0

0 1 0

0 0 1

0 0 ⫺1

80 –– 3 70 –– 3 40 –– 3

⫺–13 ↑

0

0

0

2

40 Z3 ⫽ –– 3

x2 d⫺ 3 x1

0 0 1

1 0 0

0 1 0

⫺2 ⫺1

3

2 1 3

0 10 40

P3: c Row

0

0

0

1

1

Z3 ⫽ 0

Basis x2 d⫺ 3 d⫺ 2 P3: c Row

0 0 0

517


P3: cj cB

GOAL PROGRAMMING

to substitute the values of the decision variables (x1 ⫽ 40 and x2 ⫽ 0) into the goal constraint for P4 [Eq. (11.21)] to get d⫹ 4 ⫽ 160. Thus, the cost goal is not achieved and the minimum cost of production is $160. Goal Programming Applications. Goal programming can be applied to a wide

range of problems in which the benefits are multidimensional and not directly comparable. Following is a partial list of goal programming applications: Manpower planning [26, 27] Forest management [28] Land use planning [29] Water resources [30] Marine environmental protection [31] Metal cutting [21, 23] Accounting [32] Academic planning [33, 34] Portfolio selection [35] Transportation [36–39] Marketing [40] Production planning and scheduling [41–45] Nurse scheduling [46–48] Quality control [49, 50] Additional References on Goal Programming. For additional lists of applications of goal programming, the reader is referred to the texts by Lee [51] and Schniederjan [52]. Chapter 4 of the latter text has an extensive biblio-

518


graphy (666 citations) on goal programming applications categorized by areas—accounting, agriculture, economics, engineering, finance, government, international, management, and marketing. Zanakis and Gupta [53] also has a categorized bibliographic survey of goal programming algorithms and applications. Ignizio [54] discusses advanced topics in integer and nonlinear goal programming algorithms. Zeleny [55] and Keeney and Raiffa [56] deal with other techniques of multiple-criteria optimization.

11.5

SUMMARY

In this chapter we studied optimization problems with special structures and developed algorithms for their solution. We began with integer programming problems that required some or all of the design variables to be integer valued. For the case of linear integer programs (problems with linear constraints and linear objective functions), branch-and-bound methods were developed for their solution. They involve solution of a series of linear programs that employ ‘‘continuous relaxation’’ of the original mixed-integer problem. Optimization problems with quadratic objective function and linear constraints were discussed next. Known as quadratic programming (QP) problems, these could be solved as special cases of complementary problems using Lemke’s complementary pivot method. The complementary problem basically represented the Kuhn–Tucker optimality conditions of the quadratic programs. Lemke’s algorithm is an iterative procedure for finding a solution that satisfies the Kuhn–Tucker conditions. Next, we discussed an important class of optimization problems known as goal programming. These involved problems with multiple and often conflicting objective functions. In goal programming, all the objective functions are assigned target levels for achievement and a relative priority on achieving these target levels. Goal programming then attempts to find a solution that comes ‘‘as close as possible’’ to the targets in the order of specified priorities. We also discussed an efficient algorithm called the partitioning algorithm for goal programming for solving preemptive linear goal programs.

REFERENCES 1. Watters, L., ‘‘Reduction of Integer Polynomial Programming Problems to Zero– One Linear Programming Problems,’’ Oper. R., 15(6), 1171–1174 (1967). 2. Glover, F., and E. Woolsey, ‘‘Converting 0–1 Polynomial Programming Problem to 0–1 Linear Program,’’ Oper. Res., 22, 180–182 (1974). 3. Ravindran, A., D. T. Phillips, and J. J. Solberg, Operations Research: Principles and Practice, 2nd ed., Wiley, New York, 1987. 4. Plane, D. R., and C. McMillan, Discrete Optimization, Prentice-Hall, Englewood Cliffs, NJ, 1971.

REFERENCES

519

5. Murty, K. G., Operations Research—Deterministic Optimization Models, PrenticeHall, Englewood Cliffs, NJ, 1995. 6. Rardin, R. L., Optimization in Operations Research, Prentice-Hall, Upper Saddle River, NJ, 1998. 7. Gupta, O. K., and A. Ravindran, ‘‘Branch and Bound Experiments in Convex Nonlinear Integer Programming,’’ Manag. Sci., 31(12), 1553–1546 (1985). 8. Albright, C., ‘‘Premium Solver Platform for Excel,’’ OR / MS Today, 28, 58–63 (June 2001). 9. Grossman, T. A., ‘‘Spreadsheet Add-Ins for OR / MS,’’ OR / MS Today, 29, 46–51 (Aug. 2002). 10. Saltzman, M., ‘‘Broad Selection of MIP Software Packages Available,’’ OR / MS Today, 21, 42–51 (Apr. 1994). 11. Sharda, R., Linear and Discrete Optimization and Modeling Software: A Resource Handbook, Lionheart Publishing, Atlanta, GA, 1993. 12. Nemhauser, G. L., and L. A. Woolsey, Integer and Combinatorial Optimization, Wiley, New York, 1988. 13. McCarl, B. A., H. Moskowitz, and H. Furtan, ‘‘Quadratic Programming Applications,’’ Omega, 5(1), 43–55 (1977). 14. Wolfe, P., ‘‘The Simplex Method of Quadratic Programming,’’ Econometrica, 27, 382–398 (1959). 15. Lemke, C. E., ‘‘Bimatrix Equilibrium Points and Mathematical Programming,’’ Manag. Sci., 11, 681–689 (1965). 16. Ravindran, A., and H. Lee, ‘‘Computer Experiments on Quadratic Programming Algorithms,’’ Eur. J. Oper. Res., 8(2), 166–174 (1981). 17. Cottle, R. W., and G. B. Dantzig, ‘‘Complementary Pivot Theory and Mathematical Programming,’’ J. Linear Algebra Appl., 1, 105–125 (1968). 18. Ravindran, A., ‘‘A Computer Routine for Quadratic and Linear Programming Problems,’’ Commun. ACM, 15(9), 818–820 (1972). 19. Ravindran, A., ‘‘A Comparison of Primal Simplex and Complementary Pivot Methods for Linear Programming,’’ Naval Res. Logistics Q., 20(1), 95–100 (1973). 20. Ravindran, A., ‘‘Computational Aspects of Lemke’s Algorithm Applied to Linear Programs,’’ OPSEARCH, 7(4), 241–262 (1970). 21. Philipson, R. H., and A. Ravindran, ‘‘Applications of Mathematical Programming to Metal Cutting,’’ Math. Prog. Study, 19, 116–134 (1979). 22. Armarego, E. J. A., and R. H. Brown, The Machining of Metals, Prentice-Hall, Englewood Cliffs, NJ, 1969. 23. Philipson, R. H., and A. Ravindran, ‘‘Application of Goal Programming to Machinability Data Optimization,’’ J. Mech. Design: Trans. ASME, 100, 286–291 (1978). 24. Arthur, J. L., and A. Ravindran, ‘‘An Efficient Goal Programming Algorithm Using Constraint Partitioning and Variable Elimination,’’ Manag. Sci., 24(8), 867–868 (Apr. 1978). 25. Arthur, J. L., and A. Ravindran, ‘‘PAGP—Partitioning Algorithm for (Linear) Goal Programming Problems,’’ ACM Trans. Math. Software, 6, 378–386 (1980). 26. Charnes, A., W. W. Cooper, and R. J. Niehaus, ‘‘Dynamic Multi-Attribute Models for Mixed Manpower Systems,’’ Naval Res. Logistics Q., 22(2), 205–220 (1975).

520


27. Zanakis, S. H., and M. W. Maret, ‘‘A Markovian Goal Programming Approach to Aggregate Manpower Planning,’’ 32(1), 55–63 (1981). 28. Schuler, A. T., H. H. Webster, and J. C. Meadows, ‘‘Goal Programming in Forest Management,’’ J. Forestry, 75, 320–324, June 1977. 29. Dane, C. W., N. C. Meador, and J. B. White, ‘‘Goal Programming in Land-use Planning,’’ J. Forestry, 75, 325–329, June 1977. 30. Haimes, Y. Y., W. A. Hall, and H. T. Freedman, Multiobjective Optimization in Water Resources Systems, Elsevier, Amsterdam, 1975. 31. Charnes, A., W. W. Cooper, J. Harrald, K. R. Karwan, and W. A. Wallace, ‘‘A Goal Interval Programming Model for Resource Allocation on a Marine Environmental Protection Program,’’ J. Environ. Econ. Manag., 3, 347–362 (1976). 32. Ijiri, Y., Management Goals and Accounting for Control, Rand-McNally, Chicago, IL, 1965. 33. Beilby, M. H., and T. H. Mott, Jr., ‘‘Academic Library Acquisitions Allocation Based on Multiple Collection Development Goals,’’ Computers & OR, 10(4), 335– 344 (1983). 34. Lee, S. M., and E. R. Clayton, ‘‘A Goal Programming Model for Academic Resource Allocation,’’ Manag. Sci., 17(8), 395–408 (1972). 35. Lee, S. M., and A. J. Lerro, ‘‘Optimizing the Portfolio Selection for Mutual Funds,’’ J. Finance, 28(8), 1087–1101 (1973). 36. Cook, W. D., ‘‘Goal Programming and Financial Planning Models for Highway Rehabilitation,’’ J. Oper. Res. Soc., 35(3), 217–224 (1984). 37. Lee, S. M., and L. J. Moore, ‘‘Multi-Criteria School Busing Models,’’ Manag. Sci., 23(7), 703–715 (1977). 38. Moore, L. J., B. W. Taylor, and S. M. Lee, ‘‘Analysis of a Transhipment Problem with Multiple Conflicting Objectives,’’ Comput. Oper. Res., 5, 39–46 (1978). 39. Sinha, K. C., M. Muthusubramanyam, and A. Ravindran, ‘‘Optimization Approach for Allocation of Funds for Maintenance and Preservation of the Existing Highway System,’’ Transportation Res. Record, 826, 5–8 (1981). 40. Lee, S. M., and R. Nicely, ‘‘Goal Programming for Marketing Decisions: A Case Study,’’ J. Marketing, 38(1), 24–32 (1974). 41. Arthur, J. L., and K. D. Lawrence, ‘‘Multiple Goal Production and Logistics Planning in a Chemical and Pharmaceutical Company,’’ Comput. Oper. Res., 9(2), 127– 137 (1982). 42. Deckro, R. J., J. E. Herbert, and E. P. Winkofsky, ‘‘Multiple Criteria Job-Shop Scheduling,’’ Comput. Oper. Res., 9(4), 279–285 (1984). 43. Lawrence, K. D., and J. J. Burbridge, ‘‘A Multiple Goal Linear Programming Model for Coordinated and Logistics Planning,’’ Int. J. Prod. Res., 14(2), 215– 222 (1976). 44. Lee, S. M., and J. L. Moore, ‘‘A Practical Approach to Production Scheduling,’’ Production Inventory Manag., 15(1), 79–92 (1974). 45. Lee, S. M., E. R. Clayton, and B. W. Taylor, ‘‘A Goal Programming Approach to Multi-Period Production Line Scheduling,’’ Comput. Oper. Res., 5, 205–211 (1978). 46. Arthur, J. L., and A. Ravindran, ‘‘A Multiple Objective Nurse Scheduling Model,’’ Inst. Ind. Eng. Trans., 13(1), 55–60 (1981).

PROBLEMS

521

47. Musa, A. A., and U. Saxena, ‘‘Scheduling Nurses Using Goal Programming Techniques,’’ Inst. Ind. Eng. Trans., 16(3), 216–221 (1984). 48. Trivedi, V. M., ‘‘A Mixed-Integer Goal Programming Model for Nursing Service Budgeting,’’ Oper. Res., 29(5), 1019–1034 (1981). 49. Ravindran, A., W. Shin, J. L. Arthur, and H. Moskowitz, ‘‘Nonlinear Integer Goal Programming Models for Acceptance Sampling,’’ Comput. Oper. Res., 13(6), 611– 622 (1986). 50. Sengupta, S., ‘‘A Goal Programming Approach to a Type of Quality Control Problem,’’ J. Oper. Res. Soc., 32(2), 207–212 (1981). 51. Lee, S. M., Goal Programming for Decision Analysis, Auerback Publishers, Philadelphia, PA, 1972. 52. Schniederjan, M. J., Goal Programming: Methodology and Applications, Kluwer, Norwell, MA, 1994. 53. Zanakis, S. H., and S. K. Gupta, ‘‘A Categorized Bibliographic Survey of Goal Programming,’’ Omega, 13(3), 211–222 (1985). 54. Ignizio, J. P., Goal Programming and Extensions, Health, Lexington, KY, 1976. 55. Zeleny, M., Multi Criteria Decision Making, McGraw-Hill, New York, 1982. 56. Keeney, R. L., and H. Raiffa, Decisions with Multiple Objectives: Preferences and Value Tradeofffs, Wiley, New York, 1976.

PROBLEMS

11.1. Explain under what conditions rounding off continuous solutions is a good strategy for obtaining integral values. Cite some situations where such a strategy will not work and explain why. 11.2. Explain the concepts of ‘‘branching’’ and ‘‘bounding’’ used in the branch-and-bound algorithm. 11.3. What is the meaning of ‘‘fathoming’’ a node? Under what conditions can a node be fathomed in the branch-and-bound algorithm? 11.4. Suppose we have an integer program in which a variable X1 only takes on a specified set of discrete values. Explain how to modify the branch-and-bound method to handle this variation. 11.5. Under what assumptions are the Kuhn–Tucker conditions (a) necessary and (b) sufficient for the optimality of a quadratic programming problem. 11.6. Cite conditions under which a complementary problem always has a complementary solution for all q. 11.7. Is Lemke’s algorithm guaranteed to solve a quadratic programming problem always? Why or why not?

522


11.8. What is a ‘‘ray solution’’ and what is its significance in the complementary pivot method? 11.9. What is the difference between a goal and a constraint as used in goal programming? 11.10. What are the drawbacks of using preemptive weights in goal programming? 11.11. Describe the ‘‘partitioning’’ concept and ‘‘variable elimination’’ rule in the linear goal programming algorithm. How do they help in improving the efficiency of the algorithm? 11.12. A firm has four possible sites for locating its warehouses. The cost of locating a warehouse at site i is $K. There are nine retail outlets, each of which must be supplied by at least one warehouse. It is not possible for any one site to supply all the retail outlets as shown in Figure 11.8. The problem is to determine the location of the warehouses such that the total cost is minimized. Formulate this as an integer program. 11.13. A company manufactures three products: A, B, and C. Each unit of product A requires 1 hr of engineering service, 10 hr of direct labor, and 3 lb of material. To produce one unit of product B requires 2 hr of engineering, 4 hr of direct labor, and 2 lb of material. Each unit of product C requires 1 hr of engineering, 5 hr of direct labor, and 1 lb of material. There are 100 hr of engineering, 700 hr of labor, and 400 lb of materials available. The cost of production is a nonlinear function of the quantity produced as shown below: Product B

Product A

Product C

Production (units)

Unit Cost ($)

Production (units)

Unit Cost ($)

Production (units)

Unit Cost ($)

0–40 40–100 100–150 Over 150

10 9 8 7

0–50 50–100 Over 100

6 4 3

0–100 Over 100

5 4

Figure 11.8.

PROBLEMS

523

Formulate a mixed-integer program to determine the minimum-cost production schedule. 11.14. Explain how the following conditions can be represented as linear constraints by the use of binary (0–1) integer variables: (a) Either x1 ⫹ x2 ⭐ 2 or 2x1 ⫹ 3x2 ⭓ 8. (b) Variable x3 can assume values 0, 5, 9, and 12 only. (c) If x4 ⭐ 4, then x5 ⭓ 6. Otherwise x5 ⭐ 3. (d) At least two out of the following four constraints must be satisfied: x6 ⫹ x1 ⭐ 2

x6 ⭐ 1

x7 ⭐ 5

x6 ⫹ x1 ⭓ 3

11.15. Convert the following nonlinear integer program to a linear integer program: Minimize

Z ⫽ x 21 ⫺ x1x2 ⫹ x2

Subject to x 21 ⫹ x1x2 ⭐ 8 x1 ⭐ 2

x2 ⭐ 7

x1, x2 ⭓ 0 and integer [Hint: Replace x1 by 20␦1 ⫹ 21␦2 and x2 by 20␦3 ⫹ 21␦4 ⫹ 22␦5, where ␦i 僆 (0, 1) for i ⫽ 1, 2, . . . , 5.] 11.16. Solve the following pure integer program by the branch-and-bound algorithm: Maximize

Z ⫽ 21x1 ⫹ 11x2

Subject to

7x1 ⫹ 4x2 ⫹ x3 ⫽ 13 x1, x2, x3 nonnegative integers

11.17. Solve the following mixed-integer program by the branch-and-bound algorithm: Minimize

Z ⫽ 10x1 ⫹ 9x2

Subject to x1 ⭐ 8

x2 ⭐ 10

5x1 ⫹ 3x2 ⭓ 45 x1 ⭓ 0

x2 ⭓ 0

x2 is an integer

524


11.18. A scientist has observed a certain quantity Q as a function of a variable t. She has good reasons to believe that there is a physical law relating t and Q that takes the form Q(t) ⫽ a sin t ⫹ b tan t ⫹ c

(11.24)

She wants to have the ‘‘best’’ possible idea of the value of the coefficients a, b, c using the results of her n experiments (t1, Q1; t2, Q2; t3, Q3; . . . ; tn, Qn). Taking into account that her experiments are not perfect and also perhaps that Eq. (11.24) is not rigorous, she does not expect to find a perfect fit. So she defines an error term ei ⫽ Qi ⫺ Q(ti) and has two different ideas of what ‘‘best’’ may mean: (1) Her first idea is to minimize

冘兩e 兩 n

Z1 ⫽

i

i⫽1

(2) Her second idea is to minimize

冘e n

Z2 ⫽

2 i

i⫽1

She then discovers that for physical reasons the coefficients a, b, c must be nonnegative. Taking these new constraints into consideration, she minimizes again:

冘兩e 兩 Z ⫽冘e n

(3)

Z3 ⫽

i

i⫽1 n

(4)

2 i

4

i⫽1

Show, with detailed justification, that she will have to solve, respectively: (a) A linear program in case (1). (b) A system of linear equations in case (2). [Hint: First show that in case (2) the problem reduces to minimizing an unconstrained convex function.] (c) A linear program in case (3).

PROBLEMS

525

(d) A problem of the form ␻ ⫽ Mz ⫹ q

␻⭓0

z⭓0

␻Tz ⫽ 0 in case (4)

Specify what M, z, q, ␻ are. (Hint: Show first that the problem in this case reduces to a convex quadratic program.) Note: Give all specifications you can on each problem: number of variables, number of constraints or size, etc. 11.19. A monopolist estimates the following price-versus-demand relationships for four of his products: Product 1:

x1 ⫽ ⫺1.5143p1 ⫹ 2671

Product 2:

x2 ⫽ ⫺0.0203p2 ⫹ 135

Product 3:

x3 ⫽ ⫺0.0136p3 ⫹ 0.0015p4 ⫹ 103

Product 4:

x4 ⫽ 0.0016p3 ⫺ 0.0027p4 ⫹ 19

where xj and pj are the unknown demand and price for product j. The function to be maximized is 兺4j⫽1 pj xj. The cost of production is fixed and hence can be discarded. Using the price–demand relationships, the objective function can be written as a quadratic function in either pj or xj. The products use two raw materials for production, and they give rise to the following capacity constraints: 0.026x1 ⫹ 0.8x2 ⫹ 0.306x3 ⫹ 0.245x4 ⭐ 121 0.086x1 ⫹ 0.02x2 ⫹ 0.297x3 ⫹ 0.371x4 ⭐ 250 x1, . . . , x4 ⭓ 0

p1, . . . , p4 ⭓ 0

Formulate a quadratic program to determine the optimal production levels that will maximize the total revenue 兺pj xj, and solve. 11.20. (a) Write the equivalent Kuhn–Tucker problem for the following quadratic program: Minimize

Z ⫽ x1 ⫺ x2 ⫹ x 12 ⫹ x 22 ⫺ x1x2

Subject to x1 ⫹ x2 ⫽ 1 x1, x2 ⭓ 0 (b) Show whether or not x1 ⫽ –16 , x2 ⫽ –56 in an optimal solution using the Kuhn–Tucker conditions obtained in (a).

526


11.21. The electrical resistance network shown in Figure 11.9 is required to accommodate a current flow of 10 A from junction 1 to node 4. Given the resistance of each of the arcs, the problem is to determine the equilibrium distribution of current flows that will minimize the total power loss in the network. (a) Formulate the problem as a quadratic program. (b) Write down the Kuhn–Tucker conditions. (c) Using (b) and the fact that the optimal current flows have to be positive, determine the equilibrium distribution of current flows in the network. 11.22. Consider the convex region shown in Figure 11.10, where P1, P2, P3, and P4 are the extreme points of the convex region. Let P0 ⫽ (⫺2, ⫺1) be a specified point. The problem is to determine the point in the convex region ‘‘closest’’ to point P0. (a) Formulate the above problem as a quadratic program. (Hint: Any point x in the convex region can be expressed as a convex combination of its extreme points P1, P2, P3, and P4, and the objective is to minimize 㛳x ⫺ P0㛳2.) (b) Solve using the complementary pivot method. 11.23. (a) If Q is positive semidefinite, prove that M⫽

冉

冊

Q ⫺AT A 0

is also positive semidefinite for any A matrix. (b) If Q is positive definite, what more can you say about the matrix M? 11.24. Solve the following quadratic program by the complementary pivot algorithm.

Figure 11.9. Electrical resistance network.

PROBLEMS

527

Figure 11.10.

Z ⫽ ⫺x1 ⫺ 2x2 ⫹ –21 x 12 ⫹ –21 x 22

Minimize

Subject to x1 ⫹ x2 ⭐ 4 x1, x2 ⭓ 0 Explain in two or three lines why the algorithm is guaranteed to find a solution to this problem. 11.25. Solve the following complementary problem by Lemke’s algorithm:

冢

冣

1 2 3 M⫽ 4 5 6 6 7 8

冢冣

2 q ⫽ ⫺5 ⫺3

11.26. Write down the equivalent complementary problem for the following linear programs. (In each case, state clearly what M, q, w, and z are.) (a) Maximize

Z ⫽ 2x1 ⫹ 3x2 ⫹ x3

Subject to x1 ⫺ x2 ⫹ x3 ⫽ 2 x1, x2, x3 ⭓ 0 (b) Minimize

Z ⫽ 2x1 ⫺ 3x2 ⫹ 6x3

Subject to

3x1 ⫺ 4x2 ⫺ 6x3 ⭐ 2 2x1 ⫹ x2 ⫹ 2x3 ⭓ 11 x1 ⫹ 3x2 ⫺ 2x3 ⭐ 5 x1, x2, x3 ⭓ 0

528


11.27. Prove or disprove that the complementary problem w ⫽ Mz ⫹ q, w, z ⭓ 0, wTz ⫽ 0, where

冢

冣

2 1 0 M⫽ 1 1 0 0 0 1

has a complementary solution for any q. (Give due references to any theorems/results that are applicable to your answer.) 11.28. A company has two machines that can manufacture a product. Machine 1 makes 2 units per hour, while machine 2 makes 3 per hour. The company has an order for 160 units. Each machine can be run for 40 hours on regular time; however, overtime is available. The company’s goals, in order of importance, are to: 1. Avoid falling short of meeting the demand. 2. Avoid any overtime of machine 2 beyond 10 hr. 3. Minimize the sum of overtime (assign differential weights according to the relative cost of overtime hours, assuming that the operating cost of the two machines are the same). 4. Avoid underutilization of regular working hours (assign weights according to the productivity of the machines). 11.29. A company supplies a single product from two warehouses to three of its customers. The supplies at the warehouses are 120 and 70 units, respectively. The customer demands are 70, 80, and 100 units, respectively. The net profit associated with shipping 1 unit from warehouse i to customer j is given below:

Warehouse 1 Warehouse 2

1

Customer 2

3

7 3

9 4

8 6

The company has set the following goals for shipping: 1. Meet the demand of customer 3. 2. Meet at least 60 percent of the demand of customers 1 and 2. 3. Maximize the total profit from goods shipped. 4. Balance the percentage of demands satisfied between customers 1 and 2. Formulate the above problem as a goal program. 11.30. An advertising company wishes to plan an advertising campaign in three different media—television, radio, and magazines. The purpose

PROBLEMS

529

of the advertising program is to reach as many of the potential customers as possible. Results of a market study are given below: Television Daytime Prime Time Cost of an advertising unit $40,000 Number of potential customers 400,000 reached per unit Number of women customers 300,000 reached per unit

Radio

Magazines

$75,000 900,000

$30,000 500,000

$15,000 200,000

400,000

200,000

100,000

The company has an advertising budget of $800,000 and has the following goals to achieve in order of importance: 1. At least two million exposures should take place among women. 2. Advertising on television to be limited to $500,000. 3. Achieve a total exposure of three million potential customers. 4. Advertising budget should not be exceeded by more than 10 percent. 5. The number of advertising units on radio and magazine should each be between 5 and 10. 6. At least three advertising units to be bought on daytime television and two units during prime time. Formulate the advertising problem as a goal program. 11.31. Solve the goal program formulated in problem 11.30 by the partitioning algorithm. 11.32. Solve the following goal program by the partitioning algorithm: Minimize

⫹ ⫹ ⫹ ⫺ ⫺ Z ⫽ P1d⫺ 1 ⫹ P2d2 ⫹ P3(6d3 ⫹ 5d4 ) ⫹ P4(6d4 ⫹ 5d3 )

Subject to

⫹ 5x1 ⫹ 6x2 ⫹ d⫺ 1 ⫺ d1 ⫽ 120

x2 ⫹ d2⫺ ⫺ d⫹ 2 ⫽ 11 ⫹ x1 ⫹ d⫺ 3 ⫺ d3 ⫽ 8

10x2 ⫹ d4⫺ ⫺ d⫹ 4 ⫽ 80 ⫹ xi, d⫺ i , di ⭌ 0 for all i

12 COMPARISON OF CONSTRAINED OPTIMIZATION METHODS Elsewhere in this text we examine methods from primarily two perspectives. First, some methods are included because of their historical importance (or as tools to demonstrate fundamental concepts) such as Cauchy’s method. Second, we include numerous methods that we feel are of practical importance in a modern computational environment. As we discuss these methods, we include to the extent possible, remarks that delimit their advantages and disadvantages. It is, of course, nearly impossible to be complete in this effort, and in addition we have avoided extensive discussion of rate and region of convergence. In this chapter we survey several comparative studies and give selected results from a few of these experiments. The word ‘‘experiment’’ is important, since painfully little of what is currently known of the performance of these methods on ‘‘real’’ problems has come from purely theoretical considerations. Finally, a representative list of codes is given, along with information concerning availability.

12.1

SOFTWARE AVAILABILITY

Optimization software is much more available today than in 1983, when the first edition of this text was published. Commercial computer codes, which address linear programs, integer programs, and nonlinear programs are available from a number of computer manufacturers and software companies. These software packages vary widely in complexity, ease of use, and cost [1]. Nash [2] gives a survey of NLP software that will run on PC compatibles, Macintosh systems, and UNIX-based systems. Microsoft Excel contains general-purpose optimizers, which address small optimization problems of 530


12.2

A COMPARISON PHILOSOPHY

531

almost every type. Frontline Systems [3, 4] has developed an LP solver based on the simplex algorithm and an NLP solver based on the GRG method, which are available in Excel. Modeling languages now exist that allow users to conveniently express an optimization model in a compact algebraic form, with classes of constraints and variables defined over large index sets. Models with thousands of constraints and variables can be compactly defined in a syntax that is very close to standard algebraic notation. GAMS (General Algebraic Modeling System) [5] and AMPL (A Mathematical Programming Language) [6, 7] are now widely used to solve practical problems and to test new algorithms. MATLAB is a popular software environment that offers a variety of solvers in its Optimization Toolbox [8, 9] and an interactive environment to define models, gather data, and display results. The Optimization Technology Center (OTC) [10], a cooperative activity of Northwestern University, Argonne National Laboratory, and the U.S. Department of Energy, provides a rich mix of services for the optimization user, including links to software, test results, and research in the field. The OTC website provides interactive demos, an optimization tree showing problem types and solution methods available, a software guide, an online service for posting and downloading relevant publications, a listing of seminars, and a search function. The optimization software guide on the NEOS server provides data on 129 software packages covering every optimization formulation. In this list are many well-known and commercial quality packages, including MINOS, MATLAB Optimization Toolbox, Harwell Library, and SPEAKEASY. The OTC also provides a test set of large-scale optimization problems, COPS [11, 12], which can be used by investigators to benchmark software. The American Society of Mechanical Engineers (ASME) Design Automation Committee (DAC) has now held annual conferences on optimization and related topics for over 30 years. Several of these papers have been published in the Journal of Mechanical Design. Ragsdell [13] gave a review of DAC papers up to 1977, and Mayne and Ragsdell [14] edited a group of selected papers published by the ASME in 1981. This volume contains a number of applications of optimization methods, including mechanism design, structural design, hydraulic networks, design of helical springs, and hydrostatic journal bearings. Ragsdell [15] proposed the establishment of an optimization testing center in 1981. Quite happily such centers have been established, and much more is known about the relative merits of methods which address the NLP than ever before. On the other hand, the need for meaningful comparative results continues. 12.2

A COMPARISON PHILOSOPHY

An important issue associated with any comparison is the selection of measures of merit. Even in the most qualitative setting, one must formulate merit

532

COMPARISON OF CONSTRAINED OPTIMIZATION METHODS

measure that can be used as the basis of comparison. Presentation of a new algorithm or associated analysis in the literature will include performance indicators such as number of function and constraint evaluations and, less frequently, computing time. Most performance data in the literature are given in an effort to quantify algorithm speed or efficiency. The ability of a particular algorithm or software implementation to solve a wide variety of problems (i.e., to be robust) is seldom specifically considered. This is truly regrettable, since most users consider robustness and ease of use to be at least equally as important as efficiency. It is important to recognize that the comparison of optimization algorithms and software is a multicriteria decision problem that has qualitative as well as quantitative aspects. Crowder, Dembo, and Mulvey [16] list eight measures of merit in their survey of comparative results: (1) CPU time, (2) terminal point accuracy, (3) number of iterations, (4) robustness, (5) number of function evaluations, (6) user friendliness, (7) storage requirements, and (8) basic operation count. We see that 1, 3, 5, and 8 are closely related and are indicators of efficiency. Consider now the difference between algorithm and software testing. Results from both activities are of genuine interest. Recognize the difference between an algorithm and a computer program implementation of that algorithm, or software. An algorithm can be written down step by step and comparisons conducted at this level. For instance, we might make operation counts (i.e., number of function and gradient calculations or additions, subtractions, multiplications, and divisions required per iteration) or perform other calculations related to the basic logic of the algorithm as a basis of comparison. For these algorithms to be useful, we must translate them into machinereadable form, that is, we must create software. A number of details and decisions are required to effect this translation that are difficult if not impossible to describe adequately at the algorithm level. This, coupled with the desire to monitor actual performance on representative problems, is a strong motivation for software testing. The decision to test software brings with it a number of complicating factors, since machine-dependent parameters become a part of the experiment. Eason [17] has demonstrated experimentally the influence of several machine-dependent factors. He says that ‘‘the relative performance of optimization codes does depend on the computer and compiler used for testing, and this dependence is evident regardless of the performance measure.’’ Even when the tests are performed with a single machine–compiler combination, several factors may influence the relative performance of codes. For instance, machine accuracy (word length), termination accuracy level, choice of test problems, and performance measure(s) employed may very well influence single-machine tests. The difficulties associated with testing should be kept in mind during attempts to relate the performance of one algorithm or code implementation to another, especially when someone else’s experimental results are employed. On the other hand, with the current state of development of the art and science of engineering optimization, imperfect empirical data to support decisions related to code selection seem superior to none at all. With this in mind, we

12.3

BRIEF HISTORY OF CLASSICAL COMPARATIVE EXPERIMENTS

533

survey in this chapter several classical experiments that may be useful to the reader as they consider the relative strengths and weaknesses of methods for a particular application. 12.3


Since digital computers became widely available (approximately 1960 in the United States), there have been a growing number of reports in the literature of ‘‘new’’ algorithms, often complete with supporting data suggesting superiority of the new method. In addition, a number of survey articles (see, e.g., Box [18], Brooks [19], Leon [20], and Fletcher [21]) have attempted to compare methods within classes and from class to class. A common characteristic of these surveys is the mention of a relatively small number of test problems and methods. The situation changed dramatically with the appearance of Colville’s study [22] in 1968. Colville sent eight problems having from 3 to 16 design variables, 0 to 14 constraints, and a standard timing routine to the developers of 30 codes. Each participant was invited to submit his ‘‘best effort’’ on each problem and the time required to execute the standard timing routine on his machine. The 30 codes tested represented quite a wide cross section of the available algorithms at that time. Colville placed the codes in one of five categories: search, small-step-gradient, large-step-gradient, second-derivative, and miscellaneous methods. Included in the code set are four GRG codes, several methods based on the SLP concept of Griffith and Stewart [23], a number of penalty function methods with a variety of unconstrained minimization strategies (including SUMT [24]), and a number of direct-search methods such as the complex method of Box [18]. Some codes required analytical expressions of the objective and constraint gradients, others employed numerical approximations to these gradients, while others did not employ gradient information at all. Colville collected, organized the results, and concluded that the large-stepgradient methods were the most robust and efficient as a class. Unfortunately, the Colville study contains at last three flaws. Eason [25] has shown that Colville’s timing routine does not adequately remove the effect of compiler and computing machine selection on code performance. Accordingly, the data collected at one site are not comparable to data collected at another. Furthermore, each participant was allowed to attack each problem as many times as he felt necessary to optimize the performance of his code. Thus another investigator (and certainly a general user) could not reasonably expect to produce similar results with the same code in the absence of the special insight that only its originator would possess. Finally, and possibly most important, no two participants reported solutions to the same accuracy. In 1974, Eason and Fenton [26] reported on a comparative study involving 20 codes and 13 problems. Eason did his testing on one machine and included in his test set the Colville problems plus several problems from mechanical

534


design. In addition, Eason tested primarily penalty-type methods. The major contribution of Eason’s study is the introduction of error curves as given in Figure 12.1. These data are obtained by storing intermediate iterations along the path to termination for each code. In this way a level of error is chosen as a basis of comparison after the experiments are complete. It is difficult to overestimate the value of this idea, since it makes it feasible to compare fairly the performance of codes from initial to terminal points. In effect, Eason removed the major flaws from the Colville experimental procedure. The major shortcomings of the Eason study are the lack of difficulty in the problem set and the failure to include the most powerful methods of the day. In 1977 Eric Sandgren completed a study [27, 28] of the leading NLP methods of the day. Here we examine the experimental procedure employed and the major results reported. Sandgren used the Colville and Eason problems as well as a set of geometric programming (GP) problems given by R. S. Dembo [29]. These problems provide a valuable link with previous comparisons, since they are well known and had been widely used. The GP problems are especially useful because of the availability of the powerful GP theory, which can be used to characterize the solution. In addition, this class of problems has a reputation of being difficult to solve using general-purpose NLP methods. The Sandgren test problems consist of the 13 Eason problems; problems 2, 7, and 8 of Colville’s study; the 8 Dembo problems; the welded beam problem, and 6 industrial design applications (all numbered in the order given). A complete description of these problems along with Fortran listings are given by Sandgren [27]. The problems vary in size and structure: the number of variables range from 2 to 48, the number of constraints range from 0 to 19, and the number of variable bounds range from 3 to 72. Unfortunately, only six prob-

Figure 12.1 Error curves.

12.3


535

lems contain equality constraints. The final test problem set contains 23 problems. 12.3.1

Preliminary and Final Results

The ability to solve a large number of problems in a reasonable amount of time was Sandgren’s comparison criterion, so he ranked the codes on the basis of the number of problems solved within a series of specified limits on relative solution time. The limits are based on a fraction of the average time for all codes on each problem. Furthermore, each solution time for a problem was normalized by dividing by the average solution time on that problem. This produces a low normalized solution time for a code with a relatively fast solution time and high normalized solution time for a code with a relatively slow solution time. This normalization process allows direct comparison of results on problems of different levels of difficulty. Examine the performance of the final test codes at an ␧t level of l0⫺4 as given in Table 12.1. From

Table 12.1

Code Number 1 3 8 9 10 11 12 13 14 15 16 19 20 21 22 26 27 28 29 31 32 33 34 35 a

Number of Problems Solved at ␧t ⫽ 10⫺4

Name and / or Source BIAS SEEK3 GRGDFP RALP GRG OPT GREG COMPUTE II COMPUTE II COMPUTE II COMPUTE II EXPEN (3) EXPEN (4) EXPEN (5) EXPEN (6) IPENAL (4) IPENAL (5) SUMT IV (1) SUMT IV (2) SUMT IV (4) E04HAF (0) E04HAF (1) E04HAF (2) COMET

(0) (1) (2) (3)

Times the average solution time.

Number of Problems Solved a

0.25

0.50

0.75

1.00

1.50

2.50

0 0 10 12 15 16 14 2 2 4 1 0 1 7 2 0 2 0 0 0 0 0 0 0

9 1 15 13 17 21 18 7 4 9 5 0 4 13 4 1 7 0 0 1 0 2 1 2

14 2 17 13 19 21 20 9 6 11 7 2 10 14 7 4 11 0 0 3 0 4 2 5

17 3 17 14 19 21 23 11 6 15 8 3 10 16 8 6 14 0 1 5 4 7 4 7

19 6 18 14 19 21 23 13 8 15 9 7 11 16 8 9 14 3 2 9 8 13 8 9

20 10 18 16 19 21 23 15 9 15 11 11 11 16 9 10 17 9 7 13 15 15 10 15

536


the table it is easy to identify the fast codes, since they have the large numbers in the first column labeled 0.25. The first column contains the number of problems solved by each code in 25 percent of the average solution time, and so on for the other columns. Rather than a simple ranking, Sandgren gave the performance of the final code set on the final (rated) problem set in graphical form. He normalized the results with regard to problem difficulty by using fraction of average time as the abscissa. Number of problems solved (robustness) is the ranking criterion and is therefore the ordinate. Average time is the average solution time to a given termination error for all successful codes on a given problem. Figure 12.2 gives a graphical representation of the performance of the 14 codes that solved half or more of the rated test problems. All codes in this final category exhibited superior speed and robustness. The shape of an individual utility curve gives insight to a code’s character. The perfect code would solve all 23 problems in no time, which is obviously impossible. On the other hand, codes that have a utility curve with a steep slope and a high final ordinate are both fast and robust. In Figure 12.3 the effect of the termination criterion is given. We see that all codes have difficulty solving the problems when the termination criterion becomes very small. Since several code implementations of the same (or similar) algorithm were tested, one wonders if algorithm class trends can be seen in the performance data. The three algorithm classes considered are generalized reduced gradient, exterior penalty, and interior penalty. The average performance of the best three codes in each class is shown in Figure 12.4. Clearly, the generalized reduced gradient codes are the most robust and fastest tested. Sandgren [30] has shown that the trends in Figure 12.4 are indeed statistically significant. Sandgren also tested a hybrid code consisting of a method of multipliers (BIAS) to find the neighborhood of the solution and a GRG code (OPT) to sharply refine the solution. He found this hybrid method to be the fastest and most robust when compared to the other codes tested. Finally, Sandgren used

Figure 12.2 Algorithm utility (␧t ⫽ 10⫺4).

12.3


537

Figure 12.3 Algorithm utility.

the Rosen and Suzuki [31] approach to generate NLPs with quadratic objective and constraint functions. In this way he conveniently generated problems with known characteristics, including the solution. The interested reader is referred to Sandgren’s thesis [27] for additional details. In 1980 Schittkowski [32] published the results of a very extensive study of NLP codes. His experiments included 20 codes (actually more, considering minor modifications of the same code) and 180 randomly generated test problems with predetermined character and multiple starting points. There is similarity between the Schittkowski and Sandgren studies, with several differences. The most important difference is the inclusion of successive quadratic programming (SQP) methods in the Schittkowski study. Schittkowski tested 4 penalty, 11 MOM, 3 GRG, and 4 SQP codes. Space will not allow detailed examination of Schittkowski’s experimental design or extensive results. On the other hand, we here attempt to highlight the important aspects.

Figure 12.4 Average algorithm performance.

538


Codes were evaluated using the following criteria: (1) efficiency; (2) reliability; (3) global convergence; ability to solve (4) degenerate problems, (5) ill-conditioned problems, and (6) indefinite problems; sensitivity to (7) slight problem variation and (8) position of the starting point; and (9) ease of use. Schittkowski applied each code to all of the 180 randomly generated test problems, collecting the data much in the same manner as Sandgren. Each of the nine evaluation criteria contain several subcriteria; for instance, efficiency has five subcriteria. He then chose weights using Saaty’s [33] priority theory as outlined by Lootsma [34] in order to calculate a single average efficiency measure, and similarly for the other criteria. Using the same priority theory, he chose weights for each of the nine criteria to form an overall performance measure. This produces the ranking given in Table 12.2. The weights used for each criterion are listed in Table 12.3. These priorities (weights) are a subjective matter, and a particular user might find the various criteria to be of much different relative importance and therefore would choose different weights and produce a correspondingly different ranking.

Table 12.2

Schittkowski’s Code Ranking

Code Number

Code Name

Class

Rank

3 4 7 26 6 1 2 19 14 5 17 8 11 12 23 9 10 15 21 22 18 20 16 24 13 25

VF02AD GRGA GRG2(2) FCDPAK GRG2(1) OPRQP XROP ACDPAC BIAS(1) OPT GAPFPR VF01A SALQDF SALMNF NLP LPNLP SALQDR BIAS(2) FMIN(2) FMIN(3) GAPFQL FMIN(1) FUNMIN SUMT CONMIN DFP

SQP GRG GRG SQP / Robinson GRG SQP SQP MOM MOM GRG MOM MOM MOM MOM Exterior penalty MOM MOM MOM Mixed penalty Mixed penalty MOM Mixed penalty MOM Interior penalty MOM Exterior penalty

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

REFERENCES

Table 12.3

539

Performance Measure Priorities

Efficiency Reliability Global convergence Ability to solve degenerate problems Ability to solve ill-conditioned problems Ability to solve indefinite problems Sensitivity to slight problem variation Sensitivity to the position of the starting point Ease of use

0.32 0.23 0.08 0.05 0.05 0.03 0.03 0.07 0.14 1.00

Based on his very extensive tests, Schittkowski draws some interesting algorithm class conclusions and gives recommendations for software development. He concludes that with regard to efficiency (speed) the newer methods exhibit significant improvement, and he ranks the algorithms classes more or less as (1) SQP, (2) GRG, (3) MOM, and (4) other penalty methods. He noted that the performance of GRG codes such as GRGA and OPT was relatively independent of problem condition. In fact, in some cases the performance of these codes actually improved as the test problems became increasingly ill conditioned. This is probably due to the strict constraintfollowing strategy employed. This also agrees with our own experience, which suggests that GRG-based codes are relatively insensitive to problem condition and are therefore not generally in need of special problem-scaling procedures. 12.4

SUMMARY

In this chapter we have discussed the need for empirical data to discern the relative merits of the various modern NLP methods. To learn something of the nature of the various algorithms, we test software and create a noise in the data that is both troublesome and pragmatic—troublesome in that it is sometimes impossible to point to specific algorithm characteristics as causes of particular numerical results and pragmatic in that all algorithms must be coded in machine-readable form to be of value on modern engineering problems. We have given a brief history of testing and briefly examined the Sandgren and Schittkowski studies. The chapter begins and ends with an invitation to visit the OTC. REFERENCES 1. Sharda, R., Linear and Discrete Optimization and Modeling Software: A Resource Handbook, Lionheart, Atlanta, GA, 1993.

540


2. Nash, S. G. ‘‘Nonlinear Programming Software Survey,’’ OR / MS Today, 25 (June 1998). 3. Albright, C., ‘‘Premium Solver Platform for Excel,’’ OR / MS Today, 29, 58–63 (June 2001). 4. Grossman, T. A., ‘‘Spreadsheet Add-ins for Excel,’’ OR / MS Today, 29, 46–51 (Aug. 2002). 5. Brooke, A., D. Kendrick, and A. Meerans, GAMS: A User’s Guide, Scientific Press, Redwood City, CA 1988. 6. Sharda, R., and G. Rampla, ‘‘Algebraic Modeling Languages on PC’s,’’ OR / MS Today, 22, 58–63 (1995). 7. Fourer, R., R. M. Gay, and B. W. Kermighan, ‘‘A Modeling Language for Mathematical Programming,’’ Manag. Sci., 36, 519–554 (1990). 8. Brian, R., and M. Breiner, MATLAB for Engineers, Addison-Wesley, Reading, MA 1995. 9. Hutchinson, J. J., ‘‘Matlab Software Review,’’ OR/ MS Today, 23 (Oct. 1996). 10. http: / / www.ece.northwestern.edu / OTC / . 11. Dolan, E. D., J. J. Moree´, and T. S. Munson, ‘‘Benchmarking Optimization with COPS 3.0,’’ Technical Report ANC / MCS-273, Mathematics and Computer Science Division, Argonne National Laboratory, Feb. 2004. 12. Dolan, E. D., and J. J. More´. ‘‘Benchmarking Optimization Software with Performance Profiles,’’ Math. Prog., 91, 201–213 (2002). 13. Ragsdell, K. M., ‘‘Design and Automation,’’ ASME J. Math. Design, 102, 424– 429 (1980). 14. Mayne, R. W., and K. M Ragsdell, Progress in Engineering Optimization, American Society of Mechanical Engineers, New York 1981. 15. Ragsdell, K. M., ‘‘A Proposal for the Establishment of an Optimization Software Testing Center,’’ in Nonlinear Optimization-1981 (M. J. D. Powell, Ed.), 1981, p. 465. 16. Crowder, H. P., R. S. Dembo, and J. M. Mulvey, ‘‘Reporting Computational Experiments in Mathematical Programming,’’ Math. Prog., 15, 316–329 (1978). 17. Eason, E. D., ‘‘Evidence of Fundamental Difficulties in Nonlinear Optimization Code Comparisons,’’ Proc. U.S. Bur. Stands. / Math. Programming Soc. Conf. Testing Validating Algorithms Software, Boulder, CO, Jan. 1981. 18. Box, M. J., ‘‘A New Method of Constrained Optimization and a Comparison with Other Methods,’’ Computer J., 8(1), 42–51 (1965). 19. Brooks, S. H., ‘‘A Comparison of Maximum Seeking Methods,’’ J. Oper. Res., 7(4), 430–437 (1959). 20. Leon, A., ‘‘A Comparison among Eight Known Optimizing Procedures,’’ in Recent Advances in Optimization Techniques (A. Lavi and W. Vogl, Eds.), Wiley, New York, 1966, pp. 23–42. 21. Fletcher, R., ‘‘Mathematical-Programming Methods—A Critical Review,’’ in Optimum Structural Design: Theory and Applications (R. H. Gallagher and O. C. Zienkiewicz, Eds.), Wiley, New York, 1973, pp. 51–77. 22. Colville, A. R., ‘‘A Comparative Study of Nonlinear Programming Codes,’’ Tech. Rep. No. 320-2949, IBM New York Scientific Center, NY, June 1968.

REFERENCES

541

23. Griffith, F. E., and R. A. Stewart, ‘‘A Nonlinear Programming Technique for Optimization of Continuous Processing Systems,’’ Manag. Sci., 7, 379–392 (1961). 24. Mylander, W. C., R. L. Holmes, and G. P. McCormick, A Guide to SUMT- Version, Vol. 4, Research Analysis Corporation, McLean, VA RAC-P-63, 1964. 25. Eason, E. D., ‘‘Validity of Colville’s Time Standardization for Comparing Optimization Codes,’’ Paper No. 77-DET-116, presented at the ASME Des. Eng. Tech. Conf., Chicago, Sept. 1977. 26. Eason, E. D., and R. G. Fenton, ‘‘A Comparison of Numerical Optimization Methods for Engineering Design,’’ ASME J. Eng. Ind. Ser. B, 96(1), 196–200 (Feb. 1974). 27. Sandgren, E., ‘‘The Utility of Nonlinear Programming Algorithms,’’ Ph.D. dissertation, Purdue University, Ann Arbor, MI, document no. 7813115, Dec. 1977. 28. Sandgren, E., and K. M. Ragsdell, ‘‘The Utility of Nonlinear Programming Algorithms: A Comparative Studey—Parts 1 and 2,’’ ASME J. Mech. Des., 102(3), 540–551 (July 1980). 29. Dembo, R. S., ‘‘A Set of Geometric Programming Test Problems and Their Solutions,’’ Working Paper 87, Dept. Management Sci. University of Waterloo, Ontario, Canada, Dec. 1974. 30. Sandgren, E., ‘‘A Statistical Review of the Sandgren-Ragsdell Comparative Study,’’ in Mathematical Programming Society COAL Meeting Proceedings, Boulder, CO, Jan. 5–6, 1981, pp. 72–90. 31. Rosen, J. B., and S. Suzuki, ‘‘Construction of Nonlinear Programming Test Problems,’’ ACM Commun. 8(2), Feb. 1965. Schittkowski, K., Nonlinear Programming Codes: Information Tests, Performance, Lecture Notes in Economics and Mathematical Systems, Vol. 183, Springer-Verlag, New York, 1980. 32. Schittkowski, K., Nonlinear Programming Codes: Information, Tests, Performance, Lecture Notes in Economics and Mathematical Systems, Vol. 183, SpringerVerlag, New York, 1980. 33. Saaty, T. L., ‘‘A Scaling Method for Priorities in Hierarchical Structures,’’ J. Math. Psych., 15, 234–281 (1977). 34. Lootsma, F. A., ‘‘Ranking of Nonlinear Optimization Codes According to Efficiency and Robustness,’’ in Konstrucktive Methoden der finiten nichtlinearen Optimierung (L. Collatz, G. Meinardus, and W. Wetterling, Eds.), Birkhauser, Basel, Switzerland, 1980, pp. 157–158.

13 STRATEGIES FOR OPTIMIZATION STUDIES In the preceding chapters we have concentrated on the analysis of contemporary optimization algorithms in order to understand their underlying logic as well as to develop an appreciation of their strengths and weaknesses. On the basis of this study we are now sufficiently knowledgeable to select the algorithm most appropriate for a given engineering problem and to verify whether or not a point generated by the selected algorithm is in fact a solution. While this knowledge is a necessary prerequisite for the successful execution of an optimization study, it is by no means sufficient. Many additional steps and considerations are involved in conducting such a study. It is necessary to define the optimization problem itself; to prepare it for solution; to select a suitable optimization algorithm; to choose or prepare an efficient computer implementation of this algorithm; to execute a campaign of computer runs involving possible readjustments of problem and algorithm parameters; and, finally, having obtained a reliable solution, to interpret this solution in terms of the real system as well as to implement it. Given that in most engineering applications the optimization algorithm will be available as a ‘‘canned’’ program obtained from the company’s program library or purchased/leased from software vendors, the most expensive and time-consuming parts of an optimization study will be problem formulation, preparation for solution, and preliminary troubleshooting runs. Unfortunately, these tasks involve skills that are not easily categorized, analyzed, and taught. Rather, they are acquired through practice and the careful analysis of past successes and failures. In this chapter we present some guidelines and suggest via examples some alternative approaches to problem formulation, preparation, and troubleshooting. Our discussion is by no means intended as an exhaustive treatment of these topics. It simply represents a synthesis of the experiences of the authors 542


13.1

MODEL FORMULATION

543

and those of other practitioners of the art of engineering optimization that could be gleaned through informal discussions or from casual remarks scattered in the applications literature.

13.1

MODEL FORMULATION

As noted in Chapter 1, a problem suitable for the application of optimization methods consists of a performance measure, a set of independent variables, and a set of equality and inequality constraints that constitutes the model of the system under consideration. The definition and formulation of a model that represents the real system is the most important step in conducting an optimization study, because it will dictate whether the solution obtained will be physically meaningful and realizable. The process of optimization via a model can be considered as an alternate route that allows the optimum of the real system to be found without experimenting directly with the real system. As shown in Figure 13.1, the ‘‘direct’’ route is replaced by a path that involves formulation of a model, optimization of the model, and translation of the model results into physically meaningful and applicable terms. Clearly, the indirect route to system optimization inherently involves the use of a representation of the real system that is less than perfect. In formulating this approximation or model, it is necessary to select the significant aspects of the system that must be included and those of secondary importance that can be ignored; to identify the assumptions that can be made; and to choose the form in which the model will be cast, the level of model detail, and the manner in which the model will be generated. These decisions, which together constitute the model formulation task, are inherently characterized by a certain amount of arbitrariness. Equally competent engineers viewing the same real system are quite likely to develop models that are quite different in form. Yet none of these various models of the same system, no matter how detailed and complex, can be anointed as the ‘‘true’’ model. Instead, the various models can at best be ranked on the basis of how adequately they approximate the behavior of the real system over the operating range of interest. Thus models cannot be evaluated on the basis of structure or form but only on the accuracy of their predictions. A

Figure 13.1. Optimization using a model.

544

STRATEGIES FOR OPTIMIZATION STUDIES

model involving nonlinear functions is to be preferred to a model involving linear functions only if it represents the real system more accurately. Moreover, the accuracy itself can often not be assessed precisely, and hence the selection between models can involve a considerable degree of subjectivity. For instance, one model may be more accurate than a competitive formulation over one subregion but less accurate over another subregion. In this case the engineer may well choose a model with overall lower accuracy because it is more accurate in a critical subregion. In other cases, it may, in fact, not even be possible to assess in advance the accuracy of the complete model but only that of some of its constituent equations. The key points to note are that the ties between a model and the system that it is intended to represent are at best ties of plausible association. It is part of the nature of models as simplifications of reality that there are no absolute criteria by means of which competitive models can be ranked. There are always factors involved that require subjective judgment and a ‘‘feel’’ for the behavior of the real system. Thus, there always is room for reasonable and equally competent engineers to disagree about the details of alternative model formulations. As a consequence, it is essential that the model builder be thoroughly familiar with the system to be modeled, that he or she understand the engineering principles that underlie the components of the model, and, in the case of design optimization, that he or she has mastered all of the computations that are necessary to obtain a feasible design. 13.1.1

Levels of Modeling

Since model development is an activity that requires the concentrated involvement of competent engineers thoroughly familiar with the system under study, it usually is the most expensive part of an optimization study. Since model development costs increase dramatically with the level of detail of the model, it is necessary to use considerable judgment in selecting the level that is appropriate to the goals of the study and consistent with the quality of the information available about the system. It is clearly undesirable to develop a detailed dynamic model of a plant operation if an overall input–output model will suffice. Furthermore, it is useless to develop a complex model when the data available to evaluate the model coefficients are sparse and unreliable. On the other hand, since it is the model that will be optimized and not the real system, it is pointless to carry out an optimization study with a simplistic model that will not adequately approximate the true system optimum. The ideal sought after in model development is sometimes called the ‘‘principle of optimum sloppiness’’; the model should be only as detailed as necessary for the goals of the study for which it is constructed. Achievement of this goal, however, is always difficult. While it is possible through a history of successes and failures to develop an a priori feel for the appropriate level of model complexity for specific types of systems, usually it is difficult to generalize this experience to other applications. The only foolproof way to the development of a model with optimum level of sloppiness is to proceed

13.1

MODEL FORMULATION

545

in steps with model development and optimization, starting with a very simple model and concluding with the version in which the refinements to the value of the optimum attained approach the accuracy of the data incorporated in the model. This, however, is a painstaking and methodical process whose time requirements sometimes exceed the short times that the demands of a competitive business environment allow for producing results. A common recourse to meet such deadline pressures or to avoid the labor of stepwise model development is to tailor the model to the optimization technique that is currently in favor, or with which the engineer is most familiar at the time, or that was used in a previous study. For instance, the temptation often arises to formulate all problems as linear programs or geometric programs simply because the engineer feels comfortable with this type of problem structure. To be sure, it sometimes is necessary to force a problem into a particular model form. For instance, in modeling problems involving integer variables, it is essential that the problem be forced into a tractable form such as that of a network if it is to be successfully solved. The important point is to know the impact of forcing a problem to a particular model form on the quality of the solution. Also, the capabilities and limitations of available optimization methods must be kept in mind when developing the model. It is clearly inappropriate to expect to be able to solve an NLP problem as large in dimensionality as can be accommodated by commercial LP software. However, at the modeling stage such concerns should be secondary rather than dominant considerations. To illustrate the different levels of models that might be developed for a given system, we will consider at a qualitative level a study to optimize a multiproduct plant. Example 13.1

Optimization of Polymer Plant

Problem Definition. A multiproduct polymer plant consists of several serial stages of parallel processors with in-process storage between production stages, as shown in Figure 13.2. The first stage contains several nonidentical polymerization reactor trains; the second stage consists of several parallel

Figure 13.2. Conceptual diagram of process.

546


extrusion lines of different capacity; the third stage is a blending stage, consisting of five lines, each containing one or more blending units. The final processing stage is the packaging section, which provides several packaging alternatives: bulk loading, packaging into boxes, and packaging into sacks. The plant produces a number of different types of polymer grades and types and consequently is not operated on a continuous basis. To allow for product changeovers, each processing stage has associated with it a limited amount of in-process storage. Moreover, the blending stage has associated with it several large-capacity storage vessels, which are required to assure adequate inventory for preparing product blends. A decision is made to ‘‘optimize’’ the process. It is suspected that since the current product slate is substantially different from the product slate projected when the plant and its operating procedures were designed, the plant profitability could be substantially improved. Preliminary Studies: Linear Models. As a first step in the study, a simple linear input–output model is developed to investigate what the preferred product slate might be. The linear model is based on the average production rates and processing times required by each product in each processing stage, uses average ‘‘off-specification’’ production rates (those not meeting product specifications), and assumes current levels of feed availability, current market prices, and unrestricted demand. The linear model is solved using a commercial LP code, and on the basis of several case runs it is unequivocally concluded that some products in the current slate ought to be dropped and the production of others ought to be increased. Upon further study it is revealed that as a result of increasing the production rate of some of the products and deletion of others, the production of intermediate off-specification material will be increased. This material is normally recycled internally. The increased internal recycling is likely to cause bottlenecks at the extrusion stage. The initial linear model is consequently expanded to include the key internal material flows, as shown in Figure 13.3,

Figure 13.3. Process with explicit off-specification recycle.

13.1

MODEL FORMULATION

547

as well as to incorporate explicit constraints on the maximum and minimum allowable intermediate processor rates. The revised model, considerably increased in dimensionality, remains a linear model suitable for optimization by LP methods. The case studies conducted using the expanded model indicate that the capacities of the extrusion and blending stages do, in fact, severely limit the initially projected optimum product slate. It appears that some limited expansion of capacity would be desirable. Nonlinear Models. To determine the optimum investment in additional equipment, the model is revised to incorporate nonlinear cost terms for the investment and operating costs of the extra units as a function of their throughput. The extra production capacity is easily incorporated into the linear constraint set. However, as a result of the cost terms, the objective function becomes nonlinear. The resulting linearly constrained NLP problem is solved using available software. On the basis of such optimization, an appropriate capital expenditure is proposed for plant upgrading. At this stage it might be noted by team members that the selected additional capacity was all based on average production rates and the assumption that the same product sequence was used as is currently being followed. It is suggested that alternative sequences and production run sizes ought to be considered. This introduces new considerations such as batch sizes, run lengths, product sequencing, and changeover policies into the plant model. To incorporate these elements to even a limited degree, the model must be modified to include 0–1 variables, which will be used to reflect the sequencing part of the problem. Computational considerations now come into play: available MIP codes cannot accommodate nonlinearities. Consequently, each nonlinear cost term is approximated by a fixed cost and a linear cost term. The resulting mixed-integer linear program is solved, and the solution might indicate that moderate changes in the production sequence together with subdivision into shorter product runs are attractive. At this point it may be appropriate to include time periods in the planning model as well as to incorporate the regional warehouses and the distribution network in the study. In this case, the MIP model would be further expanded to accommodate these additional elements. Simulation Models. Alternatively, questions could arise of the significance of unit shutdown and start-up transients as well as off-specification production during these periods on the plant economics. Furthermore, in-process and final product inventory levels might also need to be given closer scrutiny. If transients and inventory changes do become an important concern, then a simulation model could be developed to represent the system dynamics. In this case optimization would be carried out either by case studies or by direct search over a few variables using a direct-search optimization algorithm. Finally, if it appears that fluctuations in the production rates or down-times or changeover times of some or all of the processing stages are significant, then the simulation model may need to be refined further to include Monte

548


Carlo sampling of these fluctuating variables. In this case it will be necessary to collect statistical data on the observed fluctuations, to identify the appropriate distribution functions, and to insert those stochastic features into the model. For a discussion of such a model for the above plant, the reader is invited to consult reference 1. With this type of model, direct optimization will normally be very costly in terms of computer time, and optimization will best be performed by case studies involving a few parameter changes. In summary, depending upon the depth and direction of the study, the model used in plant optimization will range from a moderate-sized linear model to a large-scale linear program, to a large-scale mixed-integer program, to a nonlinear program, to a deterministic simulation model, and finally to Monte Carlo simulation. Clearly it would be wasteful to begin the study with the most complex model if the simplest model could serve to indicate that the potential for improvement is small. Similarly, it would be meaningless to develop and to use the detailed simulation model if no data were available to evaluate the stochastic model parameters. 13.1.2

Types of Models

From the discussion of Example 13.1, it is evident that in addition to selecting the level of detail at which the system is to be modeled, the engineer also has to choose the type of model that will be employed. Three basic model types are commonly used in optimization studies: 1. Phenomenologically based equation-oriented models 2. Response surface models 3. Procedure- or simulation-oriented models In the first case, the model is composed of the basic material balance and energy balance equations, engineering design relations, and physical properties equations, all assembled as explicit sets of equations or inequalities. To be amenable to the methods discussed in this book, the equations must be real-valued functions that can be evaluated for selected values of the independent variables. In principle, the equations could involve integral or differential operators. However, in practice it is generally best to eliminate such terms by quadrature formulas or approximations so as to result in purely algebraic model functions. Most of the examples discussed in previous chapters were equation-oriented algebraic models. Because equation-oriented models describe the system behavior at the level of basic engineering principles, they normally are valid over wider ranges of system conditions than the next category of models, response surface models. In a response surface model, the entire system or its component parts are represented by approximating equations derived by fitting coefficients to se-

13.1

MODEL FORMULATION

549

lected equation forms using directly or indirectly measured system response data. Typical approximating equations are the linear correlation function and the quadratic function with or without cross-correlation terms. Such models are used if the system response is poorly understood or too complex to allow detailed modeling from basic engineering principles. Because they are in essence correlations, response surface models are normally valid only over limited ranges of the system variables. They do, however, have the advantage of simplified structure. Finally, in the third case, the basic equations describing the behavior of the system are grouped into separate modules or subroutines that represent a particular piece of equipment or the collection of activities associated with a change in the state of the system. Each of these models or subroutines is usually a self-contained entity that may involve internal numerical procedures: equation solving, integration, or logical branching procedures. Simulation or procedure-oriented models are used when evaluation of the equations is complex, involving implicity determined variables; when the selection of a logical block of the calculation procedure or appropriate equations is subject to the state of the system; and when the model must involve stochastic elements that require Monte Carlo sampling techniques. Models of this type are generally much more complex than the previous two types and usually are much more demanding of computer resources. In most cases the choice of model type will be dictated by the level of information available about the system, the level of understanding of the phenomena occurring within the system, and the inherent complexity of the system itself. If a choice is available, then equation-oriented models are to be preferred, because they are most conveniently treated by conventional NLP techniques. Although procedure-oriented models can sometimes be more efficient, because they allow specialized treatment of recurring blocks of equations, they do make it inconvenient to access values of intermediate dependent variables, to fix their values, or to impose bounds on them as well as to interchange the roles of dependent and independent variables. We illustrate the basic differences of the three types of models with the following example. Example 13.2

Flash Separator Model

System Description. An equilibrium flash separator is a device in which a multicomponent mixture is reduced in pressure and cooled, the resulting mixture is permitted to settle into liquid and vapor phases, and the two phases are separated. As shown in Figure 13.4, such a system might consist of a valve, a heat exchanger, and a tank with an inlet nozzle and two outlet nozzles. Typically, a wire mesh pad or demister is included to trap liquid droplets entrained with the exiting vapor. Given a feed mixture with specified component mole fractions zi (moles of component i per mole of mixture) and feed rate F (moles per hour), the key design parameter of the system is the ratio

550


Figure 13.4. Equilibrium flash separator.

of vapor to feed flows V/F obtained for various values of the tank temperature T and pressure P. Suppose that a model describing the relationship between these variables is desired for a feed of fixed composition. Response Surface Model. In the absence of other information about the system, one way of describing the relationship between the ratio V/F and the tank T and P is to generate a response surface model for this feed mixture over a suitable range of T and P combinations. This requires that experiments be conducted over the selected range of T, P values and that these experimental results be used to determine the coefficients of a correlating function of appropriate form. For instance, if a quadratic function of T and P is assumed, then the response surface will take the general form

冉冊

冉

V T a ⫽ c1 ⫹ (b1, b2) ⫹ (T, P)T 11 P a12 F

a21 a22

冊冉冊 T P

(13.1)

Since a21 ⫽ a12, the response surface model will involve six constants and hence will require at least six experiments to allow the determination of these constants. Clearly, the response surface will be valid only over the range T(L) ⭐ T ⭐ T(U)

P(L) ⭐ P ⭐ P(U)

covered by the experiments. Moreover, even when restricted to that range, it will be subject to a prediction error, because the actual dependence of V/F on T and P is not likely to be exactly quadratic. Equation-Oriented Model. It is known that under suitably ideal conditions the mole fractions of vapor and liquid phases in equilibrium with each other can be related via Raoult’s law [2]. That is,

13.1

551

MODEL FORMULATION

yi pi(T) ⫽ xi P

(13.2)

where pi(T) is the partial pressure of component i at temperature T. The dependence of pi on T can be approximated using a semiempirical equation of the form ln pi ⫽ Ai ⫺

Bi T ⫺ Ci

(13.3)

where Ai, Bi, Ci are constants empirically determined for each chemical component i. An equation-oriented model relating V, L, T, and P can be constructed by using Eqs. (13.2) and (13.3) as well as the principle of conservation of mass. Thus, since the input of each component i must be equal its output rate from the system, we have Fzi ⫽ Vyi ⫹ Lxi

(13.4)

Moreover, since yi and xi are mole fractions, they must sum to 1; that is,

冘x ⫽ 1 冘y ⫽ 1 i

(13.5)

i

(13.6)

i

i

Combining (13.2) and (13.3), we also have

冉

冊

yi 1 Bi ⫽ exp Ai ⫺ xi P T ⫺ Ci

(13.7)

If the feed has three components, then Eqs. (l3.4)–(l3.7) will be a system of eight equations describing the relationship between V, L, T, P and the six additional variables xi, yi. For known F, zi and specified T, P, the nonlinear equation-oriented model will serve to determine the values of the eight variables V, L, xi, and yi. For purposes of optimization, Eqs. (13.4)–(13.7) would be treated as equality constraints. Moreover, these would probably be supplemented with the inequality constraints T(L) ⭐ T ⭐ T(U) P(L) ⭐ P ⭐ P(U) V, L, xi, yi all ⭓ 0

552


Procedure-Oriented Model. The system of Eqs. (l3.4)–(13.7) does not admit direct algebraic solution for V/F. However, it is possible to reduce it to a single nonlinear equation in the ratio V/F and the variables T, P. For fixed values of T and P, this equation can then be solved using a root-finding method to determine V/F. Specifically, it can be shown (see ref. 3) that by suitable algebraic rearrangement the equation will take the form {1 ⫺ P exp[B /(T ⫺ C ) ⫺ A ]} ⫽0 冘冋1 ⫺ z(V/F){1 ⫺ P exp[B /(T ⫺ C ) ⫺ A ]}册 i

i

i

i

i

i

i

(13.8)

i

Since Ai, Bi, Ci, and zi are constants, this equation involves only T, P, and V/F. The numerical solution of Eq. (13.8) for given values of T and P can be implemented as a self-contained subroutine or procedure that we will call FLASH. Using this routine, our model of the relationship between V/F, T, P reduces to V ⫽ FLASH(T, P) F

(13.9)

Therefore, rather than treating Eqs. (13.4)–(13.7) as equality constraints and variables V, L, xi, yi, T, P all as independent variables, the ratio V/F is simply a dependent variable whose value is fixed by those of the independent variables T, P. In contrast to the response-surface model, Eq. (13.1), the procedure-oriented model will require the internal execution of a root-finding calculation to solve Eq. (13.8) each time relation (13.9) is evaluated. Note also that if for purposes of optimization it is appropriate to impose bounds on V/F, then these bounds must be imposed on the procedure FLASH. An obvious advantage of the use of procedure FLASH is that if the system in question involves several flash units, then by suitable transfer of the parameters zi, Ai, Bi, Ci appropriate to each flash unit, the same procedure can be used to model each unit. An obvious disadvantage is that the procedure has a fixed calculation structure. Consequently, if in subsequent versions of the system model it proves desirable to fix V/F and, say, to treat the temperature T as variable, then FLASH must be discarded and a new procedure written to take its place. By contrast, in the equation-oriented model, no real difficulty is posed in fixing the value of V/F and letting T be a variable: The model equations need not be altered.

13.2

PROBLEM IMPLEMENTATION

Regardless of what level and type of model is selected, the engineer has the further choices of the manner in which the model will be implemented or assembled, the devices that will be used to prepare the resulting optimization

13.2


553

problem for solution, and the computational strategies that will be used in problem execution. These choices each have a considerable impact on the overall solution efficiency as well as on the preparation time that will be required and are therefore important components of an optimization study. 13.2.1

Model Assembly

In general, a model for an optimization study can be assembled manually and then coded into suitable problem function and derivative routines, or it can be generated using computer aids of various levels of sophistication. In the case of the equation-oriented model, the simplest form of which is a linear program, the linear equations and inequalities can be written and assembled by hand and the coefficients coded into an array suitable for processing by the LP code. Alternatively, a matrix generator can be used to automatically assemble the coefficients for certain classes of constraints. For instance, constraints of the form 兺xi ⫽ 1 could easily be generated by merely defining the indices of the variables in the sum and the numerical value of the right-handside constant. The matrix generator would then generate the appropriate array entries. All commercial LP codes require that input data be in a standard MPS (Mathematical Programming System) form in which each array entry is identified by its row, column, and numerical value, with each such triplet constituting a data record. Manual generation of such data files is very tedious, and hence some form of matrix generator is commonly used. In applications involving interconnected subsystems of various types of regular structures, it is usually efficient to use equation generators. An equation generator is written for each type of subsystem so that the assembly of the model merely involves identifying the subsystems that are present and their interconnections. For example, Bethlehem Steel Corporation has recently reported on an equation generator for their LP model that optimizes plant operations [4]. Similarly, Benjamin and co-workers [5] as well as Kilkas and Hutchison [6] have developed equation generators that are designed to prepare chemical process flowsheet models for case evaluation or direct optimization. The former generates nonlinear models and the latter linear material and energy balance models. Use of nonlinear equation generators is particularly attractive in processing applications, because such applications involve recurring components—pumps, compressors, heat exchangers, and so on—whose representation in terms of equations can be developed in general form. Because of their complexity and high development costs, equation generators are certainly not justifiable for one-time optimization studies. However, if a number of studies are contemplated, each of which could employ the generator, then development is well worthwhile. Use of a generator allows model standardization, provides convenient documentation, and minimizes modelcoding errors as well as errors of omission. In the case of response surface models, the system or its components can be employed directly to generate data from which approximating equations relating the dependent and independent variables can be derived. Typically,

554


the data will be gathered manually and fed to a regression program to determine the constants of the approximating equation. This can, of course, be automated so that data logging, fitting, and optimization are all performed without manual intervention. Often more complex models of system components are used to automatically generate response surface models for subsequent optimization. For instance, Pollack and Lieder [7] report on the use of simulations of the fluidized catalytic cracker, hydrocracker, and catalytic reformer reactors of a refinery to generate linear response surface models, which are incorporated into an overall refinery LP model. The generalized automatic generation of linear as well as quadratic response surface models from detailed nonlinear models of process flowsheets is further considered by Hughes and co-workers [8, 9] as well as McLane et al. [10]. The automated generation of response surface models from Monte Carlo simulations and their use to optimize simulated system performance is reviewed by Biles and Swain [11]. Unquestionably, automated generation and updating are preferred for all applications that will require repeated solutions. Finally, simulation or procedure-oriented models can be assembled by coding them ab initio in their entirety or by making use of simulation packages or libraries. Using such systems, a model can be assembled using the building block concept: Individual modules that represent the separate system components and are provided by the simulation package are linked together by the user either through data statements or through coding. Use of such systems is continually increasing in the processing industries [12]. Specialized packages are available for the simulation of networks involving stochastic elements [13] as well as systems involving discrete or combined discrete–continuous dynamics (see ref. 14, Chap. XIII-11 for a good review). In general, simulation or procedure-oriented models tend to be complex and hence require considerable CPU time for evaluation. Use in optimization studies is thus normally limited to a modest number of independent variables. In summary, the bulk of engineering applications tend to involve either manually prepared equation-oriented models or specially prepared standalone, procedure-oriented models. Typically, automated generation of equation-oriented models is employed only with linear or mixed-integer programming models, although their use is receiving increased attention in process industry applications. Response surface models are most frequently used in conjunction with complex simulation models to avoid direct optimization with the simulation models.

13.2.2

Preparation for Solution

Once the model is formulated and its manner of assembly is selected, the problem must be prepared for the application of an appropriate optimization algorithm. Problem preparation basically involves three types of activities:

13.2


555

1. Modifications to avoid potential numerical difficulties 2. Model rearrangement to enhance solution efficiency 3. Analysis to identify likely problem solution features Eliminating Numerical Difficulties

Numerical difficulties in problem solution that result in premature termination or an execution abort can normally be traced to four common causes: poor scaling, inconsistencies between function and derivative routines, nondifferentiable model elements, and inadequate safeguards for function arguments. In most cases these situations can be recognized from a careful review of the model and can be eliminated by simple problem modifications. Scaling. Scaling refers to the relative magnitudes of the quantities that appear

in the model to be optimized. Variable scaling refers to the relative magnitudes of the variables in the problem. Constraint scaling refers to the relative magnitudes or sensitivities of the individual constraints. Ideally, all variables in the model should be defined in such a way that their likely values fall in the range 0.1–10. This will ensure that search direction vectors and quasiNewton update vectors will take on reasonable values. If some variable values at a likely estimate of the solution fall outside this range, it is a simple matter to replace those variables by new variables multiplied by suitable constants so that the new variables are well scaled. Similarly, it is straightforward to evaluate the constraints at the solution estimate and to examine their sensitivity with respect to changes in the variable values. As a further check, one can examine the elements of the matrix of constraint gradients at the estimated solution. Generally, it is best if all constraints exhibit a similar sensitivity to changes in the variable values and if the gradients of the constraints take on the same range of values. This will ensure that violations of constraints will be equally weighted and that matrix operations involving the constraint Jacobian will not give rise to numerical inaccuracies. One remedy is to multiply the constraint functions by suitable scalars so that typical constraint values fall in the range 0.1–10 and/or that constraint gradient components take on values in the range 0.1–10. The following simple example illustrates the effects of variable and constraint scaling. Example 13.3

Problem Scaling

(a) Consider the scaling of the constraint set h1(x) ⫽ x1 ⫹ x2 ⫺ 2 ⫽ 0 h2(x) ⫽ 106x1 ⫺ 0.9 ⫻ 106x2 ⫺ 105 ⫽ 0 at the point (1.1, 1.0).

556


(b) Consider the scaling of the constraint set h1(x) ⫽ x1 ⫹ 106x2 ⫺ 2 ⫻ 106 ⫽ 0 h2(x) ⫽ x1 ⫺ 0.9 ⫻ 106x2 ⫺ 105 ⫽ 0 at the point (106, 0.9). Solution. (a) At the given point, the function values are h1(1.1, 1.0) ⫽ 01.

h2(1.1, 1.0) ⫽ 105

The given point really is reasonably close to the solution of the equation set x* ⫽ (1, 1). But, because of the poor constraint scaling, the constraint sensitivities are quite disparate. Note also that the matrix of constraint coefficients

冉

1 106

1 ⫺0.9 ⫻ 106

冊

has a determinant of ⫺1.9 ⫻ 106. The situation is easily corrected by merely dividing h2(x) by 106. The result is h1(x) ⫽ x1 ⫹ x2 ⫺ 2 ⫽ 0 h⬘2(x) ⫽

h2(x) ⫽ x1 ⫺ 0.9x2 ⫺ 0.1 ⫽ 0 106

Observe that h⬘2(l.l, 1.0) ⫽ 0.1 and the Jacobian becomes

冉

1 1

1

冊

⫺0.9

with determinant ⫺1.9. (b) At the given point, the function values are h1(106, 0.9) ⫽ ⫺105

h2(106, 0.9) ⫽ 0.9 ⫻ 105

and the constraint Jacobian is

冉

1 1

106 ⫺0.9 ⫻ 106

冊

with determinant ⫺1.9 ⫻ 106. Again the given point is reasonably close to the solution point x* ⫽ (106, 1.0). Suppose variable x1 is scaled by introducing

13.2

557


the new variable y defined as y ⫽ 10⫺6x1. In terms of the new variable, the equation set becomes h1(y, x2) ⫽ 106y ⫹ 106x2 ⫺ 2 ⫻ 106 ⫽ 0 h2(y, x2) ⫽ 106y ⫺ 9 ⫻ 105x2 ⫺ 105 ⫽ 0 If both equations are further scaled by dividing them by 106, we obtain y ⫹ x2 ⫺ 2 ⫽ 0 y ⫺ 0.9x2 ⫺ 0.1 ⫽ 0 At the given point (1, 0.9), h1⬘ ⫽ ⫺0.1

h2⬘ ⫽ 0.09

From the example it is apparent that constraint and variable scaling can be viewed in terms of the row and column scaling of the associated Jacobian matrix. Moreover, scaling may have to be performed in several successive passes, since, as was the case in part (b) of Example 13.3, scaling of a variable can impose the need to scale one or more of the constraints. A problemscaling procedure that uses successive passes to arrive at a selection of scale factors was investigated by Root and Ragsdell [15]. The objective of the scaling procedure is to determine a row scaling vector r and a column scaling vector c such that the elements J⬘ij ⫽ ri Jij cj

(13.10)

where Jij is the (i, j)th element of the matrix, whose rows are the constraint gradient vectors, lie within ratios ␧1 and ␧2 of the magnitude of the ‘‘average’’ element of J. The ‘‘average’’ element z is defined by finding the element of J of largest absolute value, identifying the corresponding row i* and column j*, and calculating the average absolute value of all nonzero elements of J excluding those of row i* or column j*. Thus, if NZ is the number of nonzero elements of J excluding those in row i* or column j*, then z⫽

1 NZ

冘冘兩J 兩 ij

i⫽i* j⫽j*

In terms of this notation, the goal of scaling is thus to find r and c such that ␧2z ⭐ 兩Jij⬘兩 ⭐ ␧1z

558


for some reasonable values of ␧1 and ␧2, say 10 and 0.1, respectively. The procedure outlined below does not guarantee attainment of the above condition for all i and j but nonetheless tends to approach it. Recursive Scaling Method At a given point x0, calculate J and z. For given values of ␧1 and ␧2, execute the following steps: Step 1. For all columns j not previously scaled, calculate cj ⫽

nj z i兩Jij兩

冘

where nj is the number of nonzero elements in column j of J. If ␧2 ⭐ cj ⭐ ␧1, reset cj ⫽ 1. Otherwise, retain cj as scale factor, update the column of J using Eq. (13.10), and recalculate z. Step 2. If at least one column j was scaled in step 1, continue. Otherwise, stop. Step 3. For all rows i not previously scaled, calculate ri ⫽

冘兩J 兩 miz j

ij

where mi is the number of nonzero elements in row i of J. If ␧2 ⭐ ri ⭐ ␧1, reset ri ⫽ 1. Otherwise, retain ri as scale factor, update the row of J using Eq. (13.10), and recalculate z. Step 4. If at least one row i was scaled in step 3, go to step 1. Otherwise stop. Note that in this procedure each row and column is scaled only once. In numerical tests performed using the MOM code BIAS [15], it was shown that a specialized form of the above scaling procedure can significantly increase solution efficiency for problems solved with BIAS. These results are confirmed by other investigators who employed scaling procedures in conjunction with penalty function methods [16]. On the other hand, Lasdon and Beck [17] investigated a more elaborate scaling method in the context of GRG algorithms and observed rather mixed results: enhanced performance in some cases but deteriorated performance in other cases. These results support the widely held belief that GRG and projection-based methods are less sensitive to scaling than penalty or multiplier methods. Lasdon and Beck [17] do suggest that simple scaling procedures such as those illustrated in Example 13.3 should be employed even with GRG methods but that more elaborate automatic scaling methods remain a topic for further research. In the experience

13.2


559

of the authors, simple scaling should be a regular component of problem preparation for optimization using any and all optimization methods. Inconsistencies. Inconsistencies between model function and derivative values can be insidious because they can catastrophically mislead the search algorithm. The simplest way to check the consistency of function and gradient values is to compute gradients by function value differences and compare these to the analytical gradients generated by the derivative routines. To avoid such difficulties, many practitioners always compute gradients by numerical differencing of function values. However, generally speaking, the use of analytical gradients is computationally more efficient, particularly if care is exercised in saving values of repeatedly recurring variable groupings. Nondifferentiability. Nondifferentiable model elements arise most commonly in two forms: conditional statements that cause different expressions or model segments to be executed depending upon the values of selected variables or functions and max/min operators. Such nondifferentiable elements cause difficulties principally when gradients are evaluated numerically in the vicinity of the switching point of the conditional statement or max/min operator. If the differencing parameter is ␦, then at x0 ⫹ ␦ one expression may be evaluated while at x0 another may have been used. The gradient thus becomes unreliable. As noted in Chapter 9 (Section 9.5), max/min operators can be eliminated by replacing them with multiple inequalities. Thus,

g(x) ⬅ max{gj (x);

j ⫽ 1, . . . , J}

j

can be replaced by g(x) ⬅ y ⭓ gj (x)

j ⫽ 1, . . . , J

However, if a model involves extensive use of conditional expressions, then it is best to use an algorithm that does not rely on gradient values. Invalid Arguments. Finally, one frequently overlooked but very common cause of computer aborts of optimization runs is failure to insert safeguards against invalid function arguments. In engineering models, terms of the form [g(x)]b or log[(g(x)], where g(x) is some expression in x, occur quite frequently. If during the course of the optimization g(x) is allowed to take on negative or zero values, then a program abort will take place. To avoid this problem, constraints of the form g(x) ⭓ ␧ ⬎ 0, where ␧ is some small tolerance value, should be added to the problem formulation. Some investigators also go so far as to suggest that all occurrences of division by variables should be eliminated from model formulations so as to avoid the possible encounters of poles of the functions or their derivatives [18].

560


Enhancing Solution Efficiency

As noted in our discussion of NLP algorithms, the difficulty of nonlinear optimization problems increases in an exponential fashion with the number of variables as well as the number of equality and inequality constraints. As a rule, linear constraints present less difficulty than nonlinear constraints and inequality constraints less difficulty than equality constraints. Consequently, there is some incentive during the problem preparation phase to attempt to rearrange the model to reduce the number of constraints, particularly the nonlinear ones, and to reduce the number of variables. These objectives can be achieved through the use of function and variable transformations, the elimination of redundant constraints, and the use of equation-sequencing techniques. Function Transformation. In the present context, a function transformation is

simply any algebraic rearrangement of the function or concatenation of a given function with another. To enhance the solution efficiency of a problem, transformations are normally chosen so as to convert nonlinear constraints to linear constraints or to convert an equality to an inequality. For instance, by cross multiplication of all terms it is obviously possible to convert a nonlinear constraint of the form y⫺1g(x) ⫽ b where g(x) is linear in x, to a linear constraint g(x) ⫺ by ⫽ 0 Similarly, nonlinear product terms of the form

写x

ant n

⭐c

n

can, by taking logarithms of both sides of the inequality, be converted to a constraint that is linear in the logarithms of the variables,

冘a

nt

ln xn ⭐ ln c

n

Of course, this transformation is only applicable if all xn and the constant c are constrained to be positive. Another popular constraint transformation device is to replace an equality constraint by two inequalities of opposite direction. Thus, hk(x) ⫽ 0 is replaced by

13.2

hk(x) ⭓ 0


561

hk(x) ⭐ 0

and

This device is only useful if, from the structure of hk and the objective function, it is obvious that one or the other of the inequalities will be limiting at the optimum, so that only one inequality need be retained in the problem formulation. If only one inequality can be used, then the feasible region will have been artificially expanded, thus making it easier to generate a feasible starting pont as well as to approach the constrained solution. However, if both inequalities must be retained, then no real computational advantage will have been achieved. Any point that is strictly interior to one inequality will be infeasible with respect to the other. Moreover, at any feasible point, both inequalities must be satisfied as equations, and hence the result of the transformation is to actually increase the number of equality constraints. To avoid such difficulties, the above inequalities are sometimes relaxed by introducing a suitable tolerance ␧ ⬎ 0, hk(x) ⭐ ␧

hk(x) ⭓ ⫺␧

The tolerance parameter serves to open up the feasible region to allow excursion from the surface hk(x) ⫽ 0. At the same time, however, the solution obtained will be in error by an amount proportional to the tolerance parameter. Example 13.4

Constraint Transformation


⫺1 ƒ(x) ⫽ x⫺1 1 x3 x4

Subject to h1(x) ⫽ x12 ⫹ x22 ⫺ 1 ⫽ 0 h2(x) ⫽ x2⫺1 x3 ⫺ 2 ⫽ 0 ⫺1 / 3 g1(x) ⫽ 3 ⫺ x⫺2 ⭓0 4 x5

g2(x) ⫽ 4 ⫺ x5 ⭓ 0 xi ⭓ 0

i ⫽ 1, . . . , 5

Observe first of all that h2(x) can readily be rewritten as a linear constraint: h⬘2(x) ⫽ x3 ⫺ 2x2 ⫽ 0 Next, by taking logarithms of both sides of the inequality, ⫺1 / 3 x⫺2 ⭐3 4 x5

Constraint g1 can be converted to

562


g1⬘(x) ⫽ ln 3 ⫹ 2 ln x4 ⫹ –13 ln x5 ⭓ 0 Finally, equality h1(x) can be written as two inequalities: g3(x) ⫽ x12 ⫹ x22 ⫺ 1 ⭓ 0 g4(x) ⫽ x21 ⫹ x22 ⫺ 1 ⭐ 0 Note, however, that ƒ(x) will be minimized by making x1 and x3 (hence, from constraint h2⬘, also x2) as large as possible. Therefore, the solution will be limited by the second inequality, and the first can be deleted. Thus, with the variable definitions y1 ⫽ ln x4

y2 ⫽ ln x5

the optimization problem takes the form Minimize

⫺1 ƒ⬘ ⫽ x⫺1 1 x3 exp(y1)

Subject to h⬘2(x) ⫽ x3 ⫺ 2x2 ⫽ 0 g1⬘(x) ⫽ ln 3 ⫹ 2y1 ⫹ –31 y2 ⭓ 0 g2⬘(x) ⫽ ln 4 ⫺ y2 ⭓ 0 g4(x) ⫽ 1 ⫺ x21 ⫺ x22 ⭓ 0 x1, x2, x3 ⭓ 0 where y1 and y2 are unrestricted in sign. Of course, since minimizing ƒ⬘ is equivalent to minimizing ln ƒ⬘, we can also replace ƒ⬘ by the separable function ln ƒ⬘ ⫽ ⫺ln x1 ⫺ ln x3 ⫹ y1 The problem of minimizing ln ƒ⬘ subject to the above transformed constraints is a convex programming problem with only one nonlinear inequality (g4) and three linear constraints. The original problem is nonconvex with three nonlinear constraints. Undoubtedly, the transformed problem will be much easier to solve than the original. Variable Transformation. A second device for reformulating a problem to en-

hance its solution efficiency is to use variable transformation. As illustrated in part in Example 13.4, variable transformations can serve to eliminate or to simplify constraints. For instance, the transformation

13.2


563

xi ⫽ y2i can be used to replace a variable xi and its nonnegativity condition xi ⭓ 0. Similarly, a constraint of the form

冘x ⫽ 1 i

with xi ⭓ 0

i

can be eliminated by defining yi ⫽

冘

xi xi

and expressing all remaining occurrences of xi by yi. The use of various clever transformations of this type is discussed in more detail by Box [19]. However, as pointed out by Sisser [20], variable transformations must be used with some care because of the following complications they can induce: 1. The occurrence of extraneous local optima 2. The degradation of convexity 3. The inhibition of convergence The following examples from Sisser [20] illustrate some of these adverse effects of poorly chosen variable transformations. Example 13.5

Variable Transformation Difficulties

(a) Consider the problem, Minimize

ƒ(x) ⫽ (x1 ⫹ 1)2 ⫹ (x2 ⫺ 4)2 x1, x2 ⭓ 0

under the transformation xi ⫽ (y2i ⫺ 1)2. (b) Consider the problem Minimize

ƒ(x) ⫽ 3x21 ⫺ 2x1x2 ⫹ 3x22

Subject to x1, x2 ⭓ 0 under the transformation xi ⫽ y2i .

564


Solution. (a) The result of the substitution is F(y) ⫽ (y41 ⫺ 2y12 ⫹ 2)2 ⫹ (y24 ⫺ 2y22 ⫺ 3)2 Since ⵜF ⫽ [2(y41 ⫺ 2y21 ⫹ 2)(4y31 ⫺ 4y1), 2(y42 ⫺ 2y22 ⫺ 3)(4y32 ⫺ 4y2)]

and ⭸2F ⫽ 32y21(y21 ⫺ 1)2 ⫹ 8(y14 ⫺ 2y12 ⫹ 2)(3y12 ⫺ 1) ⭸y21 ⭸2F ⫽ 32y22(y22 ⫺ 1)2 ⫹ 8(y24 ⫺ 2y22 ⫺ 3)(3y22 ⫺ 1) ⭸y22

it is easy to show that F(y) has six local minima. Four of these, y ⫽ (Ⳳ1, Ⳳ兹3), correspond to the minimum of ƒ(x), namely x* ⫽ (0, 4); but two, y ⫽ (Ⳳ1, 0), do not correspond to local minima of ƒ(x). The transformation has clearly introduced extraneous solutions. (b) The substituted objective function is F(y) ⫽ 3y41 ⫺ 2y21y22 ⫹ 3y42 The second derivative of F is given by ⵜ2F ⫽ 4

冉

9y12 ⫺ y22 ⫺2y1y2

⫺2y1y2 9y22 ⫺ y21

冊

which is not positive definite. For instance, along the ray y2 ⫽ 3y1, the determinant of the first principal minor is zero, while that of the second principal minor is ⫺4y21. On the other hand, since the matrix of second derivatives of ƒ is constant and positive definite, ⵜ2ƒ ⫽

冉

冊

6 ⫺2 6

⫺2

the original function is convex. Thus the transformation has destroyed a desirable problem feature. Sisser [20] gives conditions that the variable transformations and the transformed problem must satisfy to ensure that the above complications are avoided. The interested reader is invited to consult reference 20 for the details.

13.2


565

From a practical point of view, the most important difficulty to be avoided is the introduction of spurious solutions. The use of monotonic one-to-one transformations will normally achieve that aim. Redundancy Elimination. Another means of enhancing solution efficiency is to delete redundant constraints from the problem formulation, since this obviously reduces the total number of problem constraints. By a redundant constraint we mean a constraint that is not required to define the boundaries of the feasible region. For instance, the inequalities

x1 ⭐ a

x1 ⭐ b

and

where b ⬎ a ⬎ 0

are both legitimate constraints. However, since the latter lies outside of the former, it is not required to define the boundary of the feasible region in x1 and can be deleted. While redundant bounds can, of course, be very easily recognized, there are no practical procedures available that can be used to identify redundant constraints in general. Procedures are available in the linearly constrained case [21]; however, the computations involved prove worthwhile only if the optimization problem must be solved more than once, for example, with several different objective functions or with a nonlinear objective function. In spite of the lack of efficient general procedures that can guarantee the identification of all redundant constraints, it is nonetheless often possible to recognize cases of redundancy merely by inspection. Recall, for instance, that in the design problem discussed in Section 9.5 we noted by inspection that inequalities (9.53), 1.2B1 1.2B1 ⭓ R2 R3 1.5B2 1.5B2 ⭓ R2 R3 1.1B3 1.1B3 ⭓ R2 R3 could be collapsed to the single inequality R3 ⭓ R2. Similarly, we observed that constraints (9.52), T1 ⭓

1.2B1 R1

and

were subsumed by constraints (9.56),

T1 ⭓

1.2B1 R2

566


T1 ⭓

1.2B1 1.2B1 ⫹3⫹ R1 R2

since all variables were required to be nonnegative. Identification and elimination of such clearly redundant constraints should always be considered in the course of preparing a problem for solution. Equation Sequencing. Finally, as discussed in Chapter 7, the dimensionality

and number of equality constraints of a problem can be substantially reduced by explicitly or implicitly solving some of the equality constraints and using them to eliminate variables. In general, this amounts to selecting a set of independent variables and sequencing the constraints to allow the constraints to be solved for the dependent variables with as little iteration as possible. While in fortuitous situations all equalities can be eliminated without any iterative solution, normally some equations will remain that cannot be solved directly for one or more of the dependent variables. In this case, two approaches are possible: The constraints that cannot be sequenced further are solved iteratively for their dependent variables each time the problem functions must be evaluated or these remaining equations are treated explicitly as equality constraints and the remaining dependent variables as independent variables. In the former case, the iterative calculations must be carried out with sufficiently tight tolerances so as not to cause errors in the search method logic. Specifically, the tolerance on equations solved iteratively as part of the problem function calculations must be tighter than the tolerance used by the search logic in satisfying explicit equality constraints. This is inherently less efficient than applying the same tolerance to all equality constraints directly as part of the normal search logic. Consequently, the strategy in which the equations that cannot be further sequenced are treated as explicit equality constraints is usually to be preferred. As noted in Section 7.1, in all cases, bounds or inequality constraints imposed on the dependent variables must be retained in the problem formulation by expressing them in terms of the chosen set of independent variables. Analysis of Problem Features

The last stage of activities before proceeding with the actual problem optimization consists of an analysis of the final problem formulation to identify, if possible, some of the structural features that will have an impact on the solution process. These features are convexity, boundedness of the feasible region, uniqueness of the solution, and existence of a feasible point. Although the concepts themselves appear abstract and largely of theoretical interest, they in fact do have important practical significance, and hence some modest attempt at identifying these features is well worthwhile.

13.2


567

Convexity. The concept of convexity has been discussed amply in previous sections and its consequences elaborated both in terms of optimality criteria and in terms of the increased reliability of the search logic. While it can be cumbersome to prove problem convexity, it generally is rather straightforward to identify formulation elements that will make the problem nonconvex. First and foremost, it is easy to check whether the problem involves nonlinear equality constraints. If at least one is present, the problem is nonconvex. If none are present, the nonlinear inequality constraints should be tested for convexity. Two types of tests can be used: one based upon the sign of the Hessian and the other that decomposes the function as a sum of convex functions. In either case, it is advisable to begin the process of testing with the simplest nonlinear inequality and to proceed to inequalities of increasing complexity. This ordering is based on the pessimistic, but also not unrealistic, supposition that the majority of engineering problems are nonconvex and can be identified as such even from the simpler inequalities (if not from the equalities). Finally, only after the constraint set is found to be convex is it worthwhile to test for the convexity of the objective function. If at the end of the analysis the problem is found to be convex, then we first of all have assurance that any point satisfying the Kuhn–Tucker conditions will be the global minimum. Second, we will be able to select from a much broader range of algorithms for solution, and we can, in general, expect that the convergence behavior of general-purpose methods will be better than in the nonconvex case. On the other hand, if the problem is identified as nonconvex, then we must prepare to face the very real issue of multiple local solutions. Boundedness. In either the convex or the nonconvex case, there is one further important characteristic that must be checked: the boundedness of the problem. In the present context, we say that a problem is bounded if the set of feasible solutions with objective function values less than that of some nominal (initial feasible) solution can be enclosed in a finite hypercube. In engineering practice, optimization studies always involve a balancing of factors and costs to achieve finite, optimum variable values. Before any design variable can become arbitrarily large or small, some technical or economic factors will prevent it from doing so. In carrying out the optimization, however, we use a model, an approximation to the real system. The model can quite easily give rise to unbounded solutions because the modeler either intentionally or unwittingly deleted some elements or limitations of the real system. Clearly, the fact that the model is based on a real system is no assurance that mathematical pathologies will not be encountered. Cases of model unboundedness can often be identified by means of a straightforward examination of the model and can always be prevented by imposing reasonable upper and lower bounds on all variables. However, in keeping with our dictum to avoid redundant constraints, it is appropriate to first check whether these additional bounds are required before automatically imposing them.

568


Although no guaranteed test of boundedness is known to the authors, two simple checks can be constructed that will usually identify the obvious cases of unboundedness. The first test, which we call the variable test, simply involves a one-at-a-time check of each variable to establish whether its value will be bounded above and below when all remaining variables are assigned fixed values (say, those of some starting point). This test can often be carried out by inspection, without actually performing any evaluation of terms, and can normally be confined to the inequality constraints. The second test, called the partial derivative test, can be employed when function gradients are available or can be estimated. From the Kuhn–Tucker conditions, stated in terms of inequality constraints only, ⵜƒ ⫽

冘u ⵜg j

j

and

uj ⭓ 0

j

it is obvious that in order to be able to express the partial derivative ⭸ƒ/ ⭸xi of the objective function with respect to variable xi as a convex combination of inequality constraint gradients ⭸gj / ⭸xi, it is necessary that there exist at least one constraint, say l, with a negative partial derivative ⭸gl / ⭸xi and at least one, say m, with a positive value, ⭸gm / ⭸xi ⬎ 0. This statement, of course, assumes that the optimum value of xi will not be unconstrained (⭸ƒ/ ⭸xi ⫽ 0) or lie on a constraint surface but rather that it will lie at the intersection of two or more constraints. This is a reasonable and conservative assumption. The boundedness test based on this observation merely consists of checking whether for each variable xi there exists a constraint with a positive partial derivative and one with a negative partial derivative. If this is not the case for some variable, then the imposition of a bound is prudent. If no negative partial derivative exists, an upper bound should be added; if no positive partial derivative exists, a lower bound should be added. For purposes of this test, we again restrict our attention to the inequality constraints and, if many are present in the formulation, initially to the ones with simplest structure. We illustrate these simple tests with the following straightforward example. Example 13.6

Application of Boundedness Tests


⫺2 ⫺3 ƒ(x) ⫽ x⫺1 1 ⫹ x2 ⫹ x3

Subject to g1(x) ⫽ x1 ⫹ x2 ⫹ x3 ⫺ 1 ⭓ 0 g2(x) ⫽ 2 ⫺ x21 ⫺ x22 ⭓ 0 g3(x) ⫽ x1x2 x3 ⫺ 2 ⭓ 0 and all xi ⭓ 0

13.2


569

Let us first apply the variable test. When x2 and x3 are fixed with values a and b, respectively, the constraints reduce as follows: g1 is equivalent to x1 ⭓ 1 ⫺ a ⫺ b

(a lower bound)

g2 is equivalent to x1 ⭐ ⫹(2 ⫺ a2)1 / 2

(an upper bound)

and x1 ⭓ 0 When x1 and x3 are similarly fixed, the constraints become g1 is equivalent to x2 ⭓ 1 ⫺ a ⫺ b

(a lower bound)

g2 is equivalent to x2 ⭐ ⫹(2 ⫺ a2)1 / 2

(an upper bound)

and, of course, x2 ⭓ 0 Finally, when x1 and x2 are fixed at a and b, respectively, we obtain g1 is equivalent to x3 ⭓ 1 ⫺ a ⫺ b g3 is equivalent to x3 ⭓

2 ab

g2 imposes no bound on x3

(a lower bound) (another lower bound)

and

x3 ⭓ 0

Thus, x1 and x2 appear to be upper and lower bounded, while x3 is only lower bounded. Prudence would suggest the addition of a reasonable upper bound on x3, say x3 ⭐ 25. In fact, without this bound, any search method applied to this problem would tend toward the limiting solution, x* ⫽ (兹2, 兹2, ⫹ ⬁). A similar conclusion would have been found by the partial derivative test. The nonnegativity conditions (as well as g1 and g3) will yield positive partial derivatives for all xi. But constraint g2 has negative partial derivatives for only x1 and x2 [ⵜg2 ⫽ (⫺2x1, ⫺2x2, 0)]. To compensate for the missing negative partial derivative with respect to x3, we should impose an upper bound on x3. If we consider the Kuhn–Tucker gradient condition for this problem,

570


冢冣冢冣冢冣冢冣 ⫺x⫺2 1 ⫺x⫺2 2 ⫺x⫺2 3

1

⫽ u1 1

1

⫺2x1 ⫹ u2 ⫺2x2

0

x2 x3

冢

1

⫹ u3 x1x3

⫹ (u4, u5, u6) 0

x1x2

0

冣

0 0 1 0 0 1

where the identity matrix corresponds to the nonnegativity conditions on the variables, it is obvious that for no finite, positive values of xi can there be a set of nonnegative multipliers ui that will satisfy the above conditions. The imposition of the upper bound on x3 adds the column (0, 0, ⫺1)T and rectifies the situation. Note that if an equality constraint involving x3 had been included in the above formulation, then we would have ignored it for purposes of our tests. An equality may or may not have served to transfer bounds imposed on the other variables to x3. For instance, h1 ⫽ x3 ⫺ x21 ⫺ x22 ⫽ 0 would serve to transfer an upper bound on x3, since, from g2, x3 ⫽ x21 ⫹ x22 ⭐ 2 However, h2 ⫽ x3 ⫹ log(x1 ⫺ x2) ⫽ 0 would not. Disregard of equality constraints is thus a conservative policy but one that simplifies the testing. The simple boundedness tests may be too awkward to apply in very complex situations such as those involving simulation models. The decision to use them, as always, rests with the user. The key point is that the possibility of model unboundedness should always be considered. Optimization software is not as yet distinguished by its diagnostic capabilities. In our experience, it often takes many wasted runs before a user is convinced that run aborts are caused by model unboundedness rather than by a coding error in the vendor software or the user-supplied routines. Solution Uniqueness. A third important consideration in problem analysis is

recognition of the possibility of nonunique solutions or multiple local minima. In general, a convex problem may have only a single optimum objective function value but may take on that value at a family of points. Thus, although convexity guarantees a global optimum, it does not guarantee a unique solution. On the other hand, for a problem to possess more than one local minimum, it must be nonconvex, but nonconvexity is not enough to guarantee the occurrence of multiple local minima. Evidently, the convexity test discussed earlier in this section does not provide all the answers. Additional analysis is desirable to determine whether the problem will have a nonunique solution and whether it is likely to have multiple local solutions. Normally, the occurrence of a nonunique solution is an artifact of the model rather than a natural feature of the real system. The most common cause for nonuniqueness is the occurrence throughout the model of a fixed combination of two or more variables. For instance, each time either x or y appears in the problem functions, they are in the fixed combination xy. In principle, the product xy could be replaced by z ⫽ xy and the problem reduced in dimen-

13.2


571

sionality by 1. However, this type of nonuniqueness generally arises because some constraints or relations involving x or y separately were omitted in the model formulation. Consequently, the appropriate remedy, having recognized the situation, is to impose the omitted constraints. Example 13.7

Nonunique Solution


ƒ2(x) ⫽ x1⫺1x2

Subject to g1(x) ⫽ x2 ⫺ x1 ⭓ 0 x1, x2 ⭓ 0 Suppose we construct the Kuhn–Tucker gradient conditions for the problem. ⫺1 Since ⌬ƒ ⫽ (⫺x2 x⫺2 1 , x1 ) and ⵜg ⫽ (⫺1, 1), we must have

冉

⫺x2 x1⫺2

x⫺1 1

冊冉冊 ⫽ u1

⫺1

1

冉冊

⫹ (u2, u3)

1 0 0 1

At a constrained minimum, g1 must be tight; therefore, x2 ⫽ x1, and the above condition reduces to

冉冊冉冊 ⫺x⫺1 1

x⫺1 1

⫽u

⫺1

1

冉冊

⫹ (u2, u3)

1 0 0 1

Note that any positive value of x1 will satisfy the Kuhn–Tucker conditions with u1 ⫽ x⫺1 1 and u2 ⫽ u3 ⫽ 0. The minimum value of the objective function is ƒ(x) ⫽ 1, and that value will be attained at all points such that x⫺1 1 x2 ⫽ 1, that is, x2 ⫽ x1. The occurrence of x1 and x2 in fixed combination becomes more obvious if we rewrite g1 as a quotient: g1⬘ ⫽ x1⫺1x2 ⫺ 1 ⭓ 0 because then we see that these variables occur in a fixed combination in both ƒ and g1⬘. In any practical case it is highly unlikely that the designer will be indifferent to the choice of x1 ⫽ x2. Typically, some features of the real system will have been deleted from the model, perhaps, for instance, the condition x1 ⫹ x2 ⫽ 2. This additional constraint in which x1 and x2 occur separately will force the unique solution x1 ⫽ x2 ⫽ 1. Unfortunately, there are no convenient, guaranteed procedures for establishing the existence of nonunique solutions. In the case of both prototype and generalized geometric programming problems, it is sufficient to check

572


the rank of the matrix of variable exponents [22, 23]. If the rank is less than N, the number of primal variables, then nonuniqueness is assured. In such a case, it is also easy to construct a variable grouping that can be removed from the formulation [22]. This is a powerful test but does require that the problem be recast to GP form. By way of illustration, Rijckaert and Martens [24] used the exponent matrix test to show that a well-known process optimization benchmark problem in use for some 15 years had a nonunique solution. Thus, by reformulating the problem to a geometric program, it was possible to explain the number of apparently different solutions with similar objective function values that had been reported by various researchers. Since the effort involved in recasting a problem as a GP and performing a rank test is nontrivial, it can in many cases be avoided by merely inspecting the formulation to verify whether certain variable groupings tend to occur throughout the problem functions. Only if grouping is suspected is it prudent to proceed with the GP constructions. Again, the key point is to recognize that nonuniqueness is an artifact of models that can happen in practice. If it could not be identified prior to solution, it should be suspected to be occurring if multiple optimization runs from different starting points yield termination points that have similar objective function values but considerably different values of some of the variables. As noted earlier, although multiple local minima can occur whenever a nonconvex problem is solved, nonconvexity itself does not guarantee multiple local minima. In fact, it is generally difficult to prove that multiple local minima do exist except by actually identifying these points. Because the threat of local solutions is ever-present in engineering applications, while their actual existence is difficult to establish, there is a tendency to ignore this issue in application studies. In point of fact, the issue is real and must always be addressed unless it can be shown that local optima cannot exist. The problem need not be particularly complex to admit multiple local solutions. Recall that in Example 8.7 the existence of one quadratic equality constraint and two quadratic inequalities in an otherwise linear problem was sufficient to produce three distinct Kuhn–Tucker points. As shown in the following example, drawn from reference 25, multiple local minima can arise in even small problems. Example 13.8

Multiple Minima

Consider the simplified network with two heat exchangers shown in Figure 13.5. In this system, one stream with flow rate F1 is to be heated from 100⬚F to 300⬚F by exchanging heat with two hot streams with flow rates F2 and F3, available at 600 and 900⬚F, respectively. Assume that all three flows are equal to 104 lb/hr; all stream heat capacities are equal to 1 Btu/(lb 䡠 ⬚F); both overall heat transfer coefficients U1 and U2 are equal to 200 Btu/(hr 䡠 ft2 䡠 ⬚F); and exchanger costs are given by

13.2


573

Figure 13.5. Two-exchanger network, Example 13.8.

Cost ⫽ 35(area, ft2)0.6 The objective is to select the unknown temperatures T1, T2, and T3 so that the combined cost of the two exchangers is minimized. Formulation. An elementary model of a heat exchanger consists of an energy balance on each stream relating the heat transferred Q to the stream temperature difference and a design equation relating the heat transfer area A to the heat transferred and the temperature differences between the streams. The system model will, therefore, consist of the following equations: Exchanger 1 Energy balances: Design equation:

Q1 ⫽ 104(1)(T3 ⫺ 100) Q1 ⫽ 104(1)(600 ⫺ T1) Q1 ⫽ U1A1LMTD1 ⫽ 200A1LMTD1

where LMTD1 ⫽ Exchanger 2 Energy balances: Design equation:

(600 ⫺ T3) ⫹ (T1 ⫺ 100) 2

Q2 ⫽ 104(1)(300 ⫺ T3) Q2 ⫽ 104(1)(900 ⫺ T2) Q2 ⫽ U2 A2LMTD2 ⫽ 200A2LMTD2

where LMTD2 ⫽

(900 ⫺ 300) ⫹ (T2 ⫺ T3) 2

574


Note that in this case the usual log mean driving force LMTD is replaced by the arithmetic mean driving force, since the flows and heat capacities of the contacting streams are equal. Because of the simple structure of the equations as well as to compress the problem for purposes of the subsequent analysis, we will select T3 as independent design variable and reduce the problem as follows. First, the exchanger 1 energy balances can be used to relate T1 to T3, and then the result can be used in the design equation to express the area of exchanger 1 in terms of T3. Thus, A1 ⫽

50(T3 ⫺ 100) 600 ⫺ T3

Similarly, using the exchanger 2 energy balances to express T2 as a function of T3, the exchanger 2 design equation can be used to solve for A2. As a result, the area of the second exchanger is given by A2 ⫽ 50

300 ⫺ T3 600

Since the objective function is 0.6 35A0.6 1 ⫹ 35A2

the problem becomes

冋

册

50(T3 ⫺ 100) 600 ⫺ T3

Minimize

ƒ(T3) ⫽ 35

Subject to

100 ⭐ T3 ⭐ 300

0.6

⫹ 35

冋

册

50(300 ⫺ T3) 600

0.6

Analysis. At T3 ⫽ 100⬚F, ƒ(T3) ⫽ 189.3, while at T3 ⫽ 300⬚F, ƒ(T3) ⫽ 286.9. The plot of ƒ(T3) over the range 100⬚F ⭐ T3 ⭐ 300⬚F is shown in Figure 13.6. Clearly, the function has two local minima, one at T3 ⫽ 100⬚F and one at T3 ⫽ 300⬚F, and one maximum, at about T3 ⫽ 280⬚F. Presumably, any search method initiated with a temperature estimate T3 ⬎ 280⬚F will converge to the higher local minimum, while initiation with an estimate T3 ⬍ 280⬚F will converge to the lower local minimum, which is the global minimum. Note from the plot that ƒ(T3) is a concave function. In addition to illustrating a case of multiple local solutions, Example 13.8 also offers a clue that can be used in identifying situations involving local solutions. Note that the objective function contains variables raised to a power less than 1. Expressions of the form xb, where b ⬍ 1, are concave functions in x. Such expressions occur quite frequently in process optimization, because

13.2


575

Figure 13.6. Plot of Example 13.8 cost function.

the capital cost of process equipment is typically correlated in terms of power functions ax b, where x is some measure of equipment capacity [26]. The exponent b of such correlations is typically less than 1, because the cost of equipment per unit of capacity decreases as the capacity of the equipment increases. This is another instance of the phenomenon of economy of scale. Since such cost terms introduce concave elements into the objective function, we can reasonably expect that problems involving such terms will give rise to multiple local solutions. The reader should note, however, that this is a rule of thumb, not a guarantee. Recall that in Section 9.5.3 we considered a design optimization problem that had as objective function the total capital cost of the plant equipment. In that problem, all the exponents were less than 1; however, it can be shown that the constraint structure is such that only one local, hence global, minimum exists [27]. In summary, before proceeding with the problem solution, it is expedient to analyze the problem for the possibilities of nonunique solutions and multiple local optima. To check for the former situation, the formulation ought to be reviewed to identify possible occurrence of fixed-variable combinations. To check for the latter, the problem is first tested for nonconvexity. If the problem is nonconvex, then the presence of power law objective function terms should be taken as a special indication of the likely occurrence of local solutions. Feasibility. As the final element of problem presolution analysis, the formulation should be tested for feasibility. While the preceding stages of problem preparation and analysis may have resulted in a numerically stable, bounded, and nonredundant formulation, it is always possible that, as a result of poor data or calculation errors, the problem constraints simply exclude all possible

576


solutions. Thus, whether or not the optimization algorithm selected for use requires a feasible starting point, it is good practice to devote some effort to generate a feasible starting point. Obviously, if no feasible starting point can be obtained, there is little point in proceeding with optimization. Instead, the model must be inspected again and validated in a piecemeal fashion until the sources of error are identified. If a feasible point can be generated, and if the variable values at the generated point (or points) appear reasonable, then one can proceed with problem solution with a fair degree of confidence that the optimization runs will be productive. There are three procedures in common use for generating feasible starting points: random sampling, minimization of unconstrained penalty-type functions, and sequential constraint minimization. As discussed in Section 7.1.2, generation of random starting points simply involves the use of Eq. (7.1) and a pseudo-random-number generation routine to obtain trial points that are then tested for feasibility. Infeasible points are rejected, and the sampling is repeated until a feasible point is found. Unfortunately, even for moderate-sized nonlinear inequality-constrained problems (say, 15 variables and 10 inequalities), excessive computation time may be required (for instance, in excess of 500 CPU seconds on a CDC 6500 [28]. Consequently, for higher dimensionality and tightly constrained problems, the two other strategies are often preferable. A very common way of generating feasible starting points is direct minimization of the constraint infeasibilities. Under this strategy the phase I procedure consists of solving an unconstrained minimization problem whose objective function is an exterior penalty function. The starting point is thus obtained as the solution of the problem Minimize

ƒ(x) ⬅

冘[h (x)] ⫹ 冘{min[0, g (x)]} 2

2

k

k

j

j

Clearly a feasible point is one that will result in an objective function value of zero. Hence, the unconstrained minimization will be terminated when ƒ(x) becomes sufficiently small. Generally, the minimization can be simplified if the problem is posed in equality-constraint-free form. Example 13.9

Starting Point Generation, Infeasibility Minimization

Consider the problem of Example 7.2. Generation of a feasible starting point could be accomplished by solving the unconstrained problem Minimize

2 (x1 ⫹ x2 ⫹ x3 ⫺ 1)2 ⫹ (x21x3 ⫹ x2 x23 ⫹ x⫺1 2 x1 ⫺ 2)

⫹[min(0, x1)]2 ⫹ {min[0,(–12 ⫺ x1)]}2 ⫹ [min(0, x3)]2 ⫹{min[0,(–12 ⫺ x3)]}2

or, alternatively, upon substituting x1 by 1 ⫺ x2 ⫺ x3, the problem

13.2

Minimize


577

[(1 ⫺ x2 ⫺ x3)2x3 ⫹ x2 x32 ⫹ x2⫺1(1 ⫺ x2 ⫺x3) ⫺ 2]2 ⫹[min(0, 1 ⫺ x2 ⫺ x3)]2 ⫹ [min(0, x2 ⫹ x3 ⫺ –21 )]2 ⫹[min(0, x3)]2 ⫹ [min(0, –12 ⫺ x3)]2

In either case, one of the methods discussed in Chapter 3 could be used to obtain a solution. In general, either no solution, one solution, or an infinite number of solutions can be found. In the present case, the point (0.39606, 0.20788, 0.39606) with an objective value of 9 ⫻ 10⫺10 is a solution, but an infinite number of other feasible points also exist. At this point h1(x) ⫽ 0.0, h2(x) ⫽ 3 ⫻ 10⫺3, and the variable bounds are all satisfied. Some optimization algorithms require that the starting point x0 not only be feasible but also lie strictly within the inequality constraints, that is, gj (x0) ⬎ 0 for all j. In such cases the penalty function can be modified to include a tolerance parameter ␧ ⬎ 0 as follows: Minimize

冘 [h (x)] ⫹ 冘 {min[0, g (x) ⫺ ␧]} 2

2

k

k

j

j

The effect of the tolerance parameter is to force the generation of a solution x0 such that gj (x0) ⫽ ␧ ⬎ 0. However, for sufficiently large values of ␧ there may be no feasible solution satisfying gj (x) ⫽ ␧. Thus, it is sometimes necessary to repeat the minimization with a reduced value of ␧ if for the given ␧ the objective value cannot be reduced sufficiently close to zero. The minimization of the sum of squares of the constraint infeasibilities is the most frequently used phase I construction. However, the resulting objective function is quite nonlinear, and consequently, for tightly constrained highly nonlinear problems, it can be difficult to achieve an objective function value sufficiently close to zero. A more complicated but usually more reliable alternative to direct minimization of the constraint infeasibilities is a sequential minimization procedure in which each constraint is used as the objective function of a constrained minimization problem. We state the procedure for the inequality-constrained problem. Sequential Minimization Procedure

Given the constraint set gj (x) ⭓ 0 x (L) ⭐ xi ⭐ x (U) i i and a tolerance parameter ␧ ⬎ 0.

j ⫽ 1, . . . , J i ⫽ 1, . . . , N

578


Step 1. Solve the problem Minimize

⫺g1(x)

Subject to x (L) ⭐ x ⭐ x (U) terminating at the point x (1) such that g1(x (1)) ⭓ ␧. For j ⫽ 2, . . . , J, solve the series of problems as follows: Step 2. Minimize ⫺gj (x) Subject to gm(x) ⭓ 0 m ⫽ 1, . . . , J ⫺ 1 x (L) ⭐ x ⭐ x (U) with starting point x (j⫺1), terminating at the point x (j) such that gj (x (j)) ⭓ ␧. Step 3. If gj (x) ⭓ ␧ cannot be achieved for some j, reduce ␧ and repeat the subproblem solution. If ␧ becomes less than some specified zero tolerance, terminate the procedure. Note that each subproblem involves the solution of an inequalityconstrained optimization problem with known feasible starting point. Thus the method that will be used to solve the original problem can be used to solve each of the subproblems. Once the Jth subproblem is satisfactorily solved, a feasible point will have been obtained to the original constraint set. If ␧ is selected to be too large, a subproblem solution yielding gj (x) ⭓ ␧ may not exist, and hence ␧ must be reduced. To avoid an infinite computational loop, once ␧ is reduced to a value sufficiently close to zero, the procedure must be terminated, with the conclusion that no feasible point may exist. Example 13.10

Starting Point Generation, Sequential Procedure

Consider the feasible region g1(x) ⫽ x1 ⫹ x2 ⫺ 1 ⭓ 0 g2(x) ⫽ x2 ⫺ x12 ⭓ 0 g3(x) ⫽ x1 ⫺ x22 ⭓ 0 0 ⭐ (x1, x2) ⭐ 2 and suppose the sequential procedure is used to generate a feasible point with ␧ ⫽ 0.1. The first subproblem will be

13.2


Minimize

⫺g1(x) ⫽ 1 ⫺ x1 ⫺ x2

Subject to

0 ⭐ x1 ⭐ 2

579

0 ⭐ x2 ⭐ 2 and solution can be terminated at any point for which g1(x) ⭓ 0.1. Obviously, the optimum point x (1) ⫽ (2, 2) will satisfy this requirement. The second subproblem will be Minimize

⫺g2(x) ⫽ x12 ⫺ x2

Subject to g1(x) ⫽ x1 ⫹ x2 ⫺ 1 ⭓ 0 0 ⭐ (x1, x2) ⭐ 2 The starting point x (1) ⫽ (2, 2) is clearly feasible, and the suboptimal point x (2) ⫽ (0, 1) will clearly satisfy the requirement g2(x (2)) ⭓ 0.1, as will the subproblem optimum (0, 2). The third subproblem is Minimize

⫺g3(x) ⫽ x22 ⫺ x1

Subject to g2(x) ⭓ 0 g1(x) ⭓ 0 0 ⭐ (x1, x2) ⭐ 2 Again, the starting point x (2) ⫽ (0, 1) is clearly feasible. The suboptimal point x (3) ⫽ (–12 , –12 ) satisfies g3(x (3)) ⭓ 0.1, as does the optimum (0.618, 0.382) (see Figure 13.7).


580


Observe that in each case the subproblem solution iterations would normally have been stopped well short of the subproblem optimum. Hence the computational effort required per subproblem is relatively small. Equality constraints can be accommodated in the sequential approach only if the constraint elimination device is used to express the subproblems in terms of the selected independent variables. This completes our discussion of model preparation and analysis activities. In the next section, we briefly review some of the solution execution strategies that might be considered in carrying out the actual optimization. 13.2.3

Execution Strategies

Depending upon the type of model, the properties of the formulation, and its structural features, a number of different strategies can be used to actually carry out the optimization. While direct optimization using an appropriate NLP method is always possible, in many cases it is fruitful to consider alternatives such as sequential optimization using a series of subproblems or twolevel approaches in which intermediate approximate models are employed. In addition, if multiple local minima are to be expected, then some strategy must be adopted to identify the global minimum. In this section we will very briefly review these strategies and their consequences. Strategies for Equation-Oriented Models

Equation-oriented models as well as models based on response surfaces are generally solved either directly or using a sequential approach. In the direct approach, the only real choice lies in whether the problem possesses sufficient structure to be amenable to some of the specialized optimization methods of Chapter 11 or whether general-purpose NLP algorithms must be used. Normally, specialized techniques will, by definition, be preferable to generalpurpose methods and should always be investigated if repeated solution of the same type of problem will be required. However, it may take more time and effort to set up a problem for solution as a specially structured problem than to solve it as a general NLP. For instance, if a problem can be posed as a geometric programming problem, then it ought to be solved as such. However, if the problem is to be solved only once and then put aside, then it may be more efficient of the engineer’s time to simply solve it as a general NLP. Sequential approaches are those in which the problem is solved by optimizing a series of restricted subproblems. The basic idea is to attack a problem whose direct solution may be difficult by dividing the problem variables into those whose presence make the solution difficult and those whose determination would be relatively easy if the difficult variable values were fixed. The ‘‘easy’’ variable and ‘‘difficult’’ variable subproblems are solved separately with some suitable coordinating calculations to link the two. For such a strategy to be effective, the subproblems must necessarily be easier to solve

13.2


581

than the original problem, and the subproblem coordinating calculations must be such as to assure overall convergence. A detailed discussion of such strategies is certainly beyond the scope of this text; instead, the reader is directed to reference 29. We have encountered several instances of this general type of solution approach in the previous chapters. Recall, for instance, the recursive approach to solving the pooling problem as a series of linear programs (Example 8.7); the solution of goal programming problems by the partitioning algorithm (Section 11.4); and the branch-and-bound solution of integer linear programs using LP subproblems (Section 11.1.3). Sequential approaches must be tailored to the specific problem structure and hence are worthwhile if repeated solution of a class of problems with similar structure is projected. Strategies for Simulation Models

The optimization of simulation or procedure-oriented models can be executed either directly or by using various two-level strategies. In the direct approach, the entire simulation or procedure acts as a function routine that calculates performance index and constraint values and is accessed by a suitable optimization method. If the simulation or procedure output variables are in principle continuously differentiable with respect to the input variables, then any of the gradient-based constrained or unconstrained optimization algorithms can be employed. Otherwise, direct-search methods such as the complex method or a random-search method ought to be used. As an example of the former type, Sargent [30] formulates the design calculations of a multicomponent staged distillation column as a procedure-oriented model and performs the design optimization using a gradient projection method. Gradients with respect to the independent optimization variables are calculated by numerical differencing. A similar strategy is used by Ravindran and Gupta [31] to optimize a procedure-oriented model of a transportation network using a GRG algorithm. By contrast, Reklaitis and Woods [32] optimize a nondifferentiable simulation model of a nonideal multistage compressor system by using the complex method. Direct-search optimization of simulation models using either the complex or random sampling methods is, without doubt, the most frequently used approach. Direct approaches are generally suitable for studies involving no more than five optimization variables and relatively simple inequality constraints on the optimization variables. The simulation or procedure-oriented model must be structured to allow the model to be called as a subroutine by the optimizer. Moreover, if the optimization routine requires separate calls to evaluate the objective function and the constraints, then the model may have to be executed both times. There are three very common situations that arise in the direct optimization of simulation models and can cause difficulties in execution and thus merit reiteration. These are:

582


1. Implicit constraints on dependent (internal) variables 2. Implied constraints that underlie the model itself 3. Numerical procedures used within the simulation It frequently is the case that a simulation model will be constructed assuming that certain dependent variables will be within a physically meaningful range or will satisfy some reasonable engineering conditions. Unless provisions are made in advance, the values of the independent variables selected by the optimization method may force violation of these implicit constraints and thus may either cause a program abort or lead to meaningless results. The safest way to treat such constraints is to define them as explicit constraints that will be implicit nonlinear functions of the independent variables. This will, in general, require that the simulation be executed to evaluate the constraints. A similar situation and solution will be the case with implied model constraints. Often a simulation model is valid only over certain ranges of key dependent variables. Unless these ranges are treated as explicit constraints, the optimizer may force a solution that is well outside the range of validity of the model. This can be particularly insidious if an internal variable is forced outside its valid range and values of this variable are not monitored during the execution of the study. The obvious solution, again, is to define these implied constraints as explicit constraints. Finally, as mentioned earlier, if the simulation model involves internal numerical procedures, such as integration, equation solving, or the use of quadrature approximations, then care must be exercised to ensure that the tolerances used in the numerical methods are tighter than those used by the optimizer in its numerical procedures or checks. Typical optimizer numerical procedures might be differencing to obtain gradient values, Newton iterations to ensure constraint feasibility, and the usual feasibility and optimality tests. Internal numerical calculations that are conducted with looser tolerances usually will cause the optimization algorithm to be misled and hence can cause premature termination or failure to converge properly. The occurrence of the above three situations in a simulation model and the solutions outlined to resolve the associated potential difficulties can make the resulting optimization problem very difficult for direct optimization because of excessive computer time requirements. If this is the case, then it becomes increasingly appropriate to use various forms of two-level strategies. In the two-level approach, the simulation or procedure-oriented model is used to generate a response surface model expressed in terms of the independent optimization variables, and the response surface model is optimized using a suitable optimization method. The process is repeated using the response surface updated at the previously obtained optimum point until differences between successive solutions become appropriately small. Two-level strategies differ primarily in the type of approximating function used, in the level of detail at which the response surfaces are generated, and in the optimization techniques employed.

13.2


583

Most commonly, a quadratic approximating function is used, and the function coefficients are calculated from the results of a suitable pattern of model runs positioned about the current best estimate of the optimum [33] (a star or full factorial design might, for example, be used). Since the quadratic function is well behaved, any gradient-based method can be used for the optimization. The following example describes the optimization of a simulation model using a two-level strategy. Example 13.11

Optimization via a Two-Level Approach

Problem Description. The fluid-processing facility shown in Figure 13.8 consists of two parallel lines in which fluid products are pasteurized, followed by four parallel, equal-size intermediate storage vessels and four parallel, nonidentical packaging lines. In the four packaging lines the pasteurized fluids are packaged into pint, gallon, half-gallon, and gallon containers, respectively. The plant has a 16-hr operating cycle during which a projected slate of eight products must be processed and packaged into the four container sizes. Given the design daily production requirement shown in Table 13.1, the production rates of the pasteurizer and packaging lines and the storage tank volume are to be selected such as to minimize the capital cost of the equipment. Model Formulation. As described in reference 34, the system can be represented using a SLAM simulation model [35]. In this model, the products are divided among the production queues associated with each of the pasteurizers based on processing considerations. The queue of products associ-

Figure 13.8. Fluid milk plant, Example 13.11.

584


Table 13.1

Nominal Daily Production Requirements (All Units in Total Gallons)

Product Type

Total

Pints

Quarts

Half Gallons

Gallons

3.2% Milk 3% Milk 2% Milk Skim milk Orange juice Fruit drink

20,000 5,000 10,000 15,000 20,000 5,000

2000 500 1000 1000 3000 100

5000 1000 2000 2000 5000 400

10,000 3,000 5,000 10,000 10,000 500

3000 500 2000 2000 2000 4000

ated with each processor is ordered according to largest volume product first. The events that occur during the 16-hr processing day are simulated over time using a combined discrete–continuous representation. The continuous portions of the system are represented in terms of state variables. Typically the state variables represent the quantity of current product remaining to be processed on each semicontinuous unit or the current level in each storage tank. Changes in the state variables are all described by first-order difference equations (state equations), since the processing units are assumed to either operate at constant rate or be shut down. Important decision points in the time history of the system are identified in terms of specified values of the associated state variables. These decision points or events and their consequences are summarized in Table 13.2. The key assumptions built into the model are as follows:

Table 13.2

Event Index 1 2 3 4 5 6 7 8 9

Simulation State Events

Event Definition and Consequent Action A storage tank has become full. Shut off valve inputting material from pasteurizer. Positive threshold value in a tank has been crossed (three-fourths full). This allows the packagers to begin removing product. A storage tank has become empty. Shut off valve removing material. Initiate search for tank with same material. A packager has completed its current product. Initiate search of storage tanks for its next product. A pasteurizer has completed its current product. Remove next product from queue and initiate search for empty tank or tank with same product. Check for an empty tank. Check if there is an idle packager for acceptance of product. Error detection event. Negative threshold value in a storage tank has been crossed (one-fourth empty). This allows the pasteurizers to begin refilling the tank.

13.2


585

1. No product mixing is allowed in the storage tanks. 2. Once a product begins running on a pasteurizer or packager it continues running until the production quota for the product is met. 3. Once a packager finishes processing a product, it selects the next product to process on a first-come, first-served basis. 4. A tank can be refilled once it reaches one-fourth of its volume and can commence to be emptied once it is three-fourths full. The typical CPU time required to execute one 16-hr production simulation was 16 sec on a CDC 6500. Optimization Procedure. A two-level strategy is used to optimize the design of the system modeled via the simulation. At a given estimate of the design variables, a quadratic response surface is fitted to the simulation results. Since the problem involves seven design variables—the two pasteurizer rates, the storage tank volume, and the four packager rates—a factorial design using 36 simulation runs is required to evaluate all the coefficients. The response surface thus gives Production time required ⫽ q (design variables) The approximate optimization problem thus becomes Minimize

Cost of equipment ⫽ ƒ (design variables)

Subject to q (design variables) ⭐ 16 hr and Upper and lower bounds on design variables The approximate problem is solved using a GRG algorithm. Typically, 20 sec (CDC 6500) is required for a subproblem optimization. At the resulting solution, a new response surface is generated and the process continued until successive values of the optimum cost are sufficiently close. Typical results are shown in Table 13.3. For the sake of brevity, the details of the nonlinear cost function and the numerical coefficients of the seven-variable response function are not given here. The interested reader is directed to reference 34 for the details. It should be noted that since one simulation run requires, on the average, 16 CPU seconds, optimization with respect to seven design variables is quite time consuming. The five iteration cycles shown above consumed some 3000 sec or 180 simulation runs. On balance, the use of a response surface is nonetheless effective, because with 180 runs a trial optimization using the complex search made only very modest progress.

586


Table 13.3

Iteration Number 1 2 3 4 5

Results of Optimization of the Fluid Milk Process

Pasteurizing Units, gal / hr 1

2

Five Storage Tanks, gal

6000 4000 5000 5500 5400

6000 3000 3763 400 4094

3000 15000 1000 800 800

Packaging Machines, gal / hr 1

2

3

4

1000 500 500 500 499

25000 1900 1300 1116 1033

35000 2900 2600 2450 2400

1500 1200 1000 900 900

Simulated Time, hr

System Cost, $

12.12 17.04 18.05 16.26 16.74

897,490 743,314 672,045 644,172 639,855

Two-level approaches such as that illustrated in Example 13.11 can be used at varying levels of aggregation [10]. In Example 13.11, the entire system was represented in terms of a single-response-surface model. For more complex systems, it may be appropriate to generate a separate response surface for each major subsystem or even each separate component of the system. These individual response surfaces can then be combined into an approximate overall system model, which is optimized using a suitable NLP algorithm. At the resulting point the response surfaces are all updated, and the procedure is continued until convergence is attained as in the single-response-surface case. This approach has been used in the context of chemical processing systems and can be particularly effective if the process network involves recycles or loops. It is worth noting that the two-level strategy was initially investigated as a means of accelerating the convergence of flowsheet calculations involving difficult recycles [10, 12], and only subsequently has its use been advocated in performing process optimization [8, 9]. Both linear and quadratic approximate models can be used, and the optimization is carried out most effectively using either GRG or CVM algorithms. Because typically each separate component model being approximated will involve only a few different independent optimization variables, this approach can be used successfully with large-scale optimization problems involving, say, up to 25 variables. However, if each separate response surface involves most of the optimization variables, then, because of the number of simulation runs required to generate the surfaces, the dimension of the optimization problem must be restricted to a much smaller number of variables [33].

13.2


587

The execution of this type of multiple-response-surface strategy should normally be automated to as high a degree as possible. Ideally the optimization program should do the following: 1. Control the execution of portions of the simulation model to generate data for the response surface models. 2. Direct the computations required to evaluate the response surface model coefficients. 3. Transmit the calculated coefficients to the optimization model. 4. Transfer control of the optimization to the selected optimization algorithm. 5. Recover the results of the optimization and check for convergence of the two-level procedure. While full automation of the strategy is highly desirable, it clearly can be an expensive undertaking in terms of program development cost. It is therefore most appropriate for applications that will require repeated optimization studies. Strategies for Global Optimization

The successful optimization of models that are likely to involve multiple local minima requires that some strategies be employed to provide assurance that the global minimum has been found. It is desirable to find the global minimum not only because it is the best possible solution to the problem but also because local minima can severely confound the interpretation of the results of studies investigating the effects of model parameters. If, with one set of, say, cost parameters the algorithm terminates at one local minimum while with another set of parameter values a different local minimum results, then it becomes difficult to extract the true effect of the parameter changes on the solution, because much of the observed change will be due to merely having obtained a different local optimum. This difficulty is, of course, obviated if in all cases the global optimum is found. Strategies for finding global optima are still very much the subject of intensive research. A good review of the state of the art is given by Archetti and Szego [36]. The available strategies can be divided into deterministic and stochastically based approaches, Both categories include both heuristic and theoretically based strategies. The simplest and most widely used approach involves multiple optimization runs, each initiated at a different starting point. Hence, it is sometimes called the multistart strategy. Under this strategy the starting points are selected using either a deterministic grid or random sampling. In the former case, the feasible region is subdivided into disjoint sectors and the optimization is performed over each sector separately. In random sampling approaches, starting points are selected simply by sampling from a uniform distribution, as discussed in Section 7.1.2. In both cases, the global

588


minimum is defined to be the local minimum with the lowest objective function value among all local minima that can be identified. Both strategies are clearly heuristic. In the former case, the division into sectors is necessarily ad hoc: There is no assurance that each sector contains only one local minimum. In the latter case, there is no assurance that the starting points are adequately enough distributed to each lie within the neighborhood of one of the local minima. Certainly, the level of confidence in the fact that all local minima have been identified increases as the number of sectors or restart points is increased. However, that assurance is obtained at the price of a large number of optimization runs. The only successful theoretically based global optimization approaches have been those addressed to specially structured problems. For instance, Westerberg and Shah [25] developed an algorithm suitable for separable, serially structured problems that employs bounds obtained using the Lagrangian function. Cabot and Francis [37], and others since, have reported procedures suitable for linearly constrained nonconvex problems. These strategies make use of the fact that all local minima must be corner points of the region and thus develop systematic enumeration procedures for finding the best extreme point. For a discussion of other approaches for structured problems, the reader should consult the paper by McCormick [38]. Except for the approaches designed for structured problems, there are no known procedures that will find the global minimum of a general problem in a finite number of problem optimization runs. At present, the only practical strategy for locating global minima is multistart with random sampling. This strategy does force the user to decide when sufficient restarts have been used to assure that the global solution has been identified. Nevertheless, despite the possibly high costs of the repeated runs, this procedure should be used with all problems that are suspected to possess multiple local minima. 13.3

SOLUTION EVALUATION

While it may seem that the optimization study is completed once the solution has been obtained, in fact the solution only serves as the basis of the most important parts of the study: solution validation and sensitivity analysis. In most studies, it is not the numerical solution itself that is most useful; rather it is the information about the state of the system in the vicinity of the solution that provides the key insights. The answers to questions such as What constraints are active at the solution? What cost terms dominate? What is the sensitivity of the solution to parameter changes? really are the key results of the entire study. The active constraints will indicate what system resources are limiting or what design elements limit the further improvement of the system. The cost term magnitudes indicate which components of the system cost should be further refined. The sensitivity of the solution to model param-

13.3

SOLUTION EVALUATION

589

eter changes will suggest what parameter estimates need to be further improved to clearly define the true solution. In this section we briefly review the considerations involved in validating the solution and exploring its sensitivity. 13.3.1

Solution Validation

Clearly, the first issue to be addressed in studying the result of an optimization run is that of the validity of the solution. In this context, a solution is valid if it is a realizable state of the system under consideration and if it is the optimum state of the system. A realizable state is simply one in which it is possible for the system to exist. Normally if the model is a reasonably good representation of the system, then it will contain the appropriate constraints and bounds that will assure that a solution is mathematically feasible if and only if it is physically feasible or realizable. However, all models have limitations, all correlations have ranges of validity, and all data have limited accuracy. Consequently, it is necessary to review the solution in the light of the model limitations to verify that the solution does not exceed these limitations. If it does, constraints and bounds must be added to ensure that these limitations are satisfied, and the optimization runs must be repeated. Once it is established that the solution is realizable, then it is appropriate to verify whether it is reasonable to expect that it is an optimum solution for the system. The concern here is not in mathematically proving that the solution satisfies the necessary and sufficient optimality criteria but in interpreting the solution and in being able to explain why it is the optimum. For a solution to be credible, it must be possible to explain the reasons for the obtained variable values on a qualitative, even intuitive, basis. Otherwise, the validity of the solution can be accepted only as an act of faith in mathematics and computers. One approach toward gaining this insight is, as suggested by Geoffrion [39], to make use of simplified auxiliary models, which can eludicate the reasons for trends unobscured by mathematical complexity. The use of simplified auxiliary models to check trends is a device widely used in engineering and applied mathematics. For instance, in design case studies, Sherwood [40] makes considerable use of simplified, easily solved approximations to explain trends in more detailed optimal design calculations. The overall methodology recommended by Geoffrion [39] consists of the following steps: 1. Reduce the model detail until it can be solved by simple algebra. 2. From the auxiliary model, derive the general trend of the optimal solution as a function of key model parameters. 3. Calculate specific predictions from the auxiliary model and test them by rerunning the full model.

590


4. If the numerical optimization results confirm the trends projected from the auxiliary model, then we have succeeded in explaining the key trade-off involved. Normally, this process might be continued with several different auxiliary models, each of which reflects certain key factors or trade-offs in the full model or system. As a result of interpretive studies of this type, we not only will have verified the solution but also will have achieved some understanding of the physical meaning behind the calculated optimal solution. Thus, we will have begun to close the gap between the model optimum and the system optimum.

13.3.2

Sensitivity Analysis

The next stage in evaluating the solution is to determine the sensitivity of the solution to changes in the model parameters or assumptions. This is called sensitivity analysis. There are a number of reasons for performing a detailed sensitivity analysis: 1. To find one or more parameters with respect to which the optimal solution is very sensitive. If such parameters exist, then it may be worthwhile to change the corresponding system features. For instance, if it is discovered that the optimal solution is very sensitive to labor availability (a constraint placed on the system), then it may be possible to increase labor availability by employing overtime labor and thus increase the overall system profitability. 2. To extract information about additions or modifications to the system so as to improve the overall operation. Thus, one can obtain information as to the advisability of adding new production capacity or increasing intermediate storage. 3. To clarify the effect on the system of variations in imprecisely known parameters. Some model parameters may be subject to considerable uncertainty. Sensitivity analysis can indicate whether it is worthwhile to expend resources to obtain better estimates of these parameter values. Alternatively, it may be that parameters that initially appeared to be critical turn out to be unimportant and hence need not be further refined. 4. To suggest the likely effects of variations in uncontrollable external parameters. Some system inputs, for example, product demands, may be outside of our control. Parameter sensitivity will give estimates of the effects of product demands on profits and hence allow the user to plan for a range of economic returns.

13.3

591

SOLUTION EVALUATION

Because this type of information is so important in implementing a solution on the real system, a detailed sensitivity analysis is, in many cases, more valuable than the actual optimal solution itself. Sensitivity information is normally extracted in two ways: through the Lagrange multiplier values and through parameter case study runs. The reader will recall that in the case of linear programming we can also easily obtain information about sensitivity with respect to the objective function coefficients without recalculating the optimum. However, that option is available only with LP models. As discussed in Chapter 5, the Lagrange multipliers (as well as the Kuhn–Tucker multipliers) are measures of the rate of change of the objective function with respect to a change in the right-hand side of the constraint. Thus, given a constraint hk(x) ⫽ bk, the multiplier vk is equal to vk ⫽

⭸ƒ ⭸bk

Similarly, for an inequality constraint gj (x) ⭓ dj, the multiplier uj is equal to uj ⫽

⭸ƒ ⭸dj

As a first-order approximation, the change of the objective function value resulting from changes in the right-hand side of constraints will be given by ƒ(x) ⫺ ƒ(x*) ⫽

冘冉⭸⭸bƒ 冊 ⌬b ⫹ 冘冉⭸⭸dƒ 冊 ⌬d ⫽ 冘v ⌬b ⫹ 冘u ⌬d k

k

k

j

j

j

k

k

k

j

j

j

(13.11) This estimate of the change in the optimal objective function value is likely to be quite good provided that the changes ⌬dj, ⌬bk are small, that the same constraints remain tight at the optimum, and that only a few constraints are perturbed at a time. In the next example we illustrate the use of Eq. (13.11) to estimate optimal objective function changes. Example 13.12

Use of Lagrange Multipliers

Consider the welded beam design problem of Example 1.2. With the parameter values specified in that example, the design optimization problem takes the form

592


Minimize

ƒ(x) ⫽ 1.10471x21x2 ⫹ 0.04811x3x4(14 ⫹ x2)

Subject to g1(x) ⫽

再

␶d 1 3(28 ⫹ x2) ⫺ ⫹ F 2x21x22 x21x2[x22 ⫹ 3(x1 ⫹ x3)2] ⫹

冎

4.5(28 ⫹ x2)2[x22 ⫹ (x1 ⫹ x3)2] x21x22(x22 ⫹ 3(x1 ⫹ x3)2)2

1/2

⭓0

where F ⫽ 6000 lb and ␶d ⫽ 13,600 psi g2(x) ⫽ x32x4 ⫺ 12.8 ⭓ 0 g3(x) ⫽ x4 ⫺ x1 ⭓ 0 g4(x) ⫽ x3x34(1 ⫺ 0.02823x3) ⫺ 0.09267 ⭓ 0 g5(x) ⫽ x1 ⫺ 0.125 ⭓ 0 g6(x) ⫽ x33x4 ⫺ 8.7808 ⭓ 0 x2 and x3 ⭓ 0 Suppose that the problem is solved using a penalty function method (SUMTIV) to obtain the solution shown in Table 13.4. Estimate the effects on the objective function value of a change in ␶d from 13,600 to 13,240 psi. Solution. From Table 13.4 it is evident that the first four constraints are tight at the optimum. To perform the sensitivity analysis, we first need to obtain values of the multipliers of the tight constraints if they are not provided automatically by the optimization routine. Thus, we must solve the system

Table 13.4

Quantity x1 x2 x3 x4

Solution of Welded Beam Problem

Starting Point 2.0 4.0 4.0 3.0

Optimum Point 0.2536 7.141 7.1044 0.2536

ƒ(x)

28.0671

2.34027

g1 g2 g3 g4 g5 g6

1.719225 35.2 1.0 95.7119 1.875 183.2192

2.327 ⫻ 1.863 ⫻ 7.341 ⫻ 3.392 ⫻ 0.1280 82.169

10⫺5 10⫺3 10⫺6 10⫺6

13.3

ⵜƒ(x*) ⫽

冘u ⵜg (x*) j

SOLUTION EVALUATION

593

for uj ⭓ 0

j

In the present case, the system reduces to

冢

冣冢冣冢冣

9.222 0.24249 0.28518 0 0 0 3.604 50.473 ⫺1 0 0 1.0 0 0 0.00977 1.0961

T

u1 4.0018 u2 0.1577 ⫽ u3 0.25797 u4 7.226

In general this system need not be square. In fact, normally there will be more equations than unknown multipliers, and hence a linear least-squares solution will have to be obtained. In the present case, however, the number of variables is exactly equal to the number of tight inequalities, and thus the Kuhn–Tucker conditions can be solved directly to yield u ⫽ (0.650553, 0.008193518, 1.997586, 4.3292586) Now, if ␶d is changed to 13,240, then the change in the constant associated with g1 is from 2.2667 to 2.20667. Since the constraint is written g1 ⫽ 2.2667 ⫺ q(x) ⭓ 0 or ⫺q(x) ⭓ ⫺2.2667

we have d1 ⫽ ⫺2.2667. Hence, ⌬d1 ⫽ (⫺2.20662) ⫺ (⫺2.2667) or ⌬d1 ⫽ 0.06. From Eq. (13.11), the predicted change in the objective function corresponding to this change in the constraint will be ⌬ƒ ⫽ u1 ⌬d1 ⫽ 0.650553(0.06) ⫽ 0.039033

Thus, the new objective function value will be ƒ(x) ⫽ 2.34027 ⫹ 0.039033 ⫽ 2.379303 By way of comparison, when the problem is resolved using SUMT with the new value of ␶d, the actual new objective function value is 2.38043. The prediction using the Lagrange multiplier is thus quite good. Note that the magnitudes of the multiplier values further indicate that the objective function value will be most sensitive with respect to the fourth constraint, the buckling constraint, and least sensitive with respect to the second constraint. While the multipliers yield useful information about constraint sensitivity, they do not indicate sensitivity with respect to individual function parameters.

594


Thus, it is expedient to execute a series of case studies in which selected model parameters are reassigned values and the problem is rerun.

13.4

SUMMARY

This chapter focused on the steps essential to the successful execution of an optimization study. The basic concepts involved in model formulation were reviewed and the levels and types of optimization models (equation oriented, response surface, and procedure oriented) as well as their relative advantages and disadvantages were considered. Next, problem implementation details were summarized: how models can be implemented or assembled, what devices can be employed to prepare the problem for optimization, and what computational strategies should be used in executing the optimization itself. Problem preparation was found to involve three types of activities: modification to avoid numerical difficulties, model rearrangement to enhance solution efficiency, and problem analysis to establish essential problem features. The topics discussed under these categories include problem scaling, function and variable transformations, elimination of redundant constraints, and equation sequencing. Simple constructions useful in identifying key problem features such as convexity, boundedness, solution uniqueness, and problem feasibility were reviewed. Then optimization strategies for both equationoriented and simulation models were considered, as were problems associated with identifying global optima. We found that while much research is under way on the latter issue, in general, the multistart strategy is presently the only practical recourse. The chapter concluded with a consideration of issues involved in evaluating the solution provided by the optimizer. We observed that quantitative as well as qualitative validation and interpretation of a solution to establish its credibility are highly desirable. Finally, the concept of sensitivity analysis was discussed, and the use of Lagrange multipliers to estimate the sensitivity of the solution to parameter changes was reviewed. We hope that this chapter has helped the reader realize that the execution of an optimization study cannot merely be reduced to the running of a canned program. It involves careful consideration of many aspects of the application problem itself, the model used to represent the application, and the algorithms employed to perform the number crunching.

REFERENCES 1. Embury, M. C., G. V. Reklaitis, and J. M. Woods, ‘‘Scheduling and Simulation of a Staged Semi-Continuous Multi-Product Process,’’ Proceedings of Pacific Chemical Engineering Congress, Denver, CO, Aug. 1977, Vol. 2.

REFERENCES

595

2. Bett, K. E., J. S. Rawlinson, and G. Saville, Thermodynamics for Chemical Engineers, MIT Press, Cambridge, MA, 1975. 3. King, J. C., Separation Processes, McGraw-Hill, New York, 1971. 4. Scribner, B. F., and D. H. Berlau, ‘‘Optimal Planning for Product / Production Facility Studies (OPPLAN),’’ Proc. Amer. Inst. Ind. Eng. Conf., San Francisco, May 1979, pp. 383–387. 5. Benjamin, D. R., M. H. Locke, and A. W. Westerberg, ‘‘ASCEND: Friendly Interactive Flowsheeting System,’’ paper 25b, presented at AIChE Summer Meeting, Detroit, Aug. 1981. 6. Kilkas, A. C., and H. P. Hutchison, ‘‘Process Optimization Using Linear Models, Computers Chem. Eng., 4, 39–48 (1980). 7. Pollack, H. W., and W. D. Lieder, ‘‘Linking Process Simulators to a Refinery Linear Programming Model’’ in Computer Applications to Chemical Engineering (R. G. Squires and G. V. Reklaitis, Eds.), ACS Symposium Series Vol. 124, American Chemical Society, Washington, DC, 1980. 8. Parker, A. L., and R. R. Hughes, ‘‘Approximation Programming of Chemical Processes: Parts 1 and 2,’’ Computers Chem. Eng., 5, 123–131 (1981). 9. Biegler, L. T., and R. R. Hughes, ‘‘Approximation Programming of Chemical Processes with Q / LAP,’’ Chem. Eng. Prog., 77, 76–83 (1981). 10. McLane, M., M. K. Sood, and G. V. Reklaitis, ‘‘A Hierarchical Strategy for Large Scale Process Calculations,’’ Computers Chem. Eng., 3, 383–394 (1979). 11. Biles, W. E., and J. J. Swain, Optimization and Industrial Experimentation, WileyInterscience, New York, 1980, Chap. 7. 12. Rosen, E. M., ‘‘Steady State Chemical Process Simulation: A State of the Art Review,’’ in Computer Applications to Chemical Engineering (R. G. Squires and G. V. Reklaitis, Eds.), ACS Symposium Series Vol. 124, American Chemical Society, Washington, DC, 1980. 13. Pritsker, A. A. B., Modeling and Analysis Using Q-GERT Networks, WileyHalsted, New York, 1977. 14. Salvendy, G. (Ed), Handbook of Industrial Engineering, Wiley-Interscience, New York, 1981. 15. Root, R. R., and K. M. Ragsdell, ‘‘Computational Enhancements of the Method of Multipliers,’’ ASME J. Mech. Des., 102, 517–523 (1980). 16. Keefer, D. L., and B. Gottfried, ‘‘Differential Constraint Scaling in Penalty Function Optimization,’’ AIIE Trans., 2, 281–286 (1970). 17. Lasdon, L. S., and P. O. Beck, ‘‘Scaling Nonlinear Programs,’’ Oper. Res. Lett., 1, 6–9 (1981). 18. Berna, T. J., M. H. Locke, and A. W. Westerberg, ‘‘A New Approach to Optimization of Chemical Processes,’’ AIChE J., 26, 37–43 (1980). 19. Box, M. J., ‘‘A Comparison of Several Current Optimization Methods and the Use of Transformations in Constrained Problems,’’ Computer J., 9, 67–77 (1966). 20. Sisser, F. S., ‘‘Elimination of Bounds in Optimization Problems by Transforming Variables,’’ Math. Prog., 20, 110–121 (1981). 21. Luenberger, D. G., Introduction to Linear and Nonlinear Programming, AddisonWesley, Reading, MA, 1973, Chap. 5.

596


22. Reklaitis, G. V., and D. J. Wilde, ‘‘A Differential Algorithm for Polynomial Programs,’’ DECHEMA Monogr., 67, 324–350 (1970). 23. Reklaitis, G. V., and D. J. Wilde, ‘‘Geometric Programming via a Primal Auxiliary Problem,’’ AIIE Trans., 6, 308–317 (1974). 24. Rijckaert, M. J., and X. M. Martens, ‘‘Analysis and Optimization of the WilliamsOtto Process by Geometric Programming,’’ AIChE J., 20, 742–750 (1974). 25. Westerberg, A. W., and J. V. Shah, ‘‘Assuring a Global Optimum by the Use of an Upper Bound on the Lower (Dual) Bound,’’ Computers Chem. Eng., 2, 83–92 (1978). 26. Guthrie, K. M., Process Plant Estimating, Evaluation, and Control, Craftsman Book Company, Solana Beach, CA, 1974. 27. Knopf, F. C., M. R. Okos, and G. V. Reklaitis, ‘‘Optimal Design of Batch / SemiContinuous Processes,’’ Ind. Eng. Chem. Proc. Design Devel., 21, 79–86 (1982). 28. Fattler, J. E., G. V. Reklaitis, Y. T. Sin, R. R. Root, and K. M. Rasgsdell, ‘‘On the Computational Utility of Polynomial Geometric Programming Solution Methods,’’ Math. Prog., 22, 163–201 (1982). 29. Geoffrion, A. M., ‘‘Generalized Benders Decomposition,’’ JOTA, 10, 237–251 (1972). 30. Sargent, R. W. H., and K. Gannibandara, ‘‘Optimum Design of Plate Distillation Columns.’’ in Optimization in Action (L. C. W. Dixon, Ed.), Academic, New York, 1976. 31. Gupta, A., and A. Ravindran, ‘‘Optimization of Transportation Networks,’’ in Stochastic Modeling of Large Scale Transportation Networks (J. J. Solberg, Ed.), Final Rep., Automotive Transportation Center, Purdue University, West Lafayette, IN, Apr. 1981. 32. Reklaitis, G. V., and J. M. Woods, ‘‘Optimization of a Non-Ideal Staged Compression Facility,’’ Proc. 1974 Purdue Compressor Technol. Conf., Purdue University, West Lafayette, IN, July 1974, pp. 415–422. 33. Biles, W. E., and J. J. Swain, Optimization and Industrial Experimentation, Wiley, New York, 1980, Chap. 7. 34. Knopf, F. C., G. V. Reklaitis, and M. R. Okos, ‘‘Combined Design and Scheduling of Regenerative Noncontinuous Processes,’’ Symp. Prod. Planning Multi-Product Plants, AIChE Nat. Meet., Houston, TX, Apr. 1981. 35. Pritsker, A. A. B., and C. P. Pegden, Introduction to Simulation and SLAM, Systems Publ. Corp. and Halsted Press, New York, 1979. 36. Archetti, F., and G. P. Szego, ‘‘Global Optimization Algorithms,’’ in Nonlinear Optimization: Theory and Algorithms (L. C. W. Dixon, E. Spedicato, and G. P. Szego, Eds.), Birkhauser, Boston, 1980. 37. Cabot, A. V., and R. L. Francis, ‘‘Solving Nonconvex Quadratic Minimization Problems by Ranking Extreme Points,’’ Oper. Res., 18, 82–86 (1970). 38. McCormick, G. P., ‘‘Attempts to Calculate Global Solutions to Problems That May Have Local Minima,’’ in Numerical Methods for Nonlinear Optimization (F. A. Lootsma, Ed.), Academic, New York, 1972. 39. Geoffrion, A. M., ‘‘The Purpose of Mathematical Programming in Insight, not Numbers,’’ Interfaces, 7, 81–92 (1976). 40. Sherwood, T., A Course in Process Design, MIT Press, Cambridge, MA, 1963.

597

PROBLEMS

PROBLEMS

13.1. Apply the sealing procedure to the two problems of Example 13.3 and compare the results to the scaling obtained in that example. 13.2. Determine row and column scale factors and the resulting scaled form of the Jacobian matrix

冤

冥

10⫺4 0 1

102 107 105

2 104 10⫺3

13.3. Determine scale factors for the constraint set g1(x) ⫽ x1 ⫹ 2x2 ⫺ x3 ⫺2 ⭓ 0 g2(x) ⫽ 100x21 ⫹ 0.02x22 ⫹ 10x23 ⭓ 0 g3(x) ⫽ 105x1 ⫹ 103x22 ⫺ 4 ⫻ 106x23 ⭓ 0 in the vicinity of the feasible point (0.01, 1, 0.01). 13.4. Consider the problem Minimize

ƒ(x) ⫽ x21 ⫹ x22

Subject to g(x) ⫽ x2 ⫺ max[2x1⫺1, –23 ⫺ –21 (x1 ⫺ 3)2] x1, x2 ⭓ 0 (a) show that ⵜg is discontinuous at the point (2, 1). (b) Reformulate constraint g(x) so that this problem is avoided. 13.5. A model for the optimization of a through circulation dryer can be formulated as follows [Chung, S. F., Can. J. Chem. Eng., 51, 262 (1973)]: Maximize

再

ƒ(x) ⫽ 0.033x1

0.036 1 ⫺ exp(⫺107.9x2 x1⫺0.41) ⫹ 0.095 ⫺

冋

⫻ ln

9.27 ⫻ 10⫺4x10.41 x2

册冎

1 ⫺ exp(⫺5.39x2 /x0.41 1 ) 1 ⫺ exp(⫺107.9x2 /x0.41 1 )

⫺1

598


Subject to g1(x) ⫽ 2 ⫻ 103 ⫺ 4.62 ⫻ 10⫺6x12.85x2 ⫺ 1.055x1 ⭓ 0 g2(x) ⫽ 175 ⫺ 9.84 ⫻ 10⫺4x1.85 1 x2 ⭓ 0 g3(x) ⫽ 2 ⫺ 109.6

再

x2 0.036 0.41 x1 1 ⫺ exp(⫺107.9x2 /x10.41) ⫹ 0.095 ⫺

冋

⫻ ln

9.27 ⫻ 10⫺4x10.41 i x2

册冎

1 ⫺ exp(⫺5.39x2 /x0.41 1 ) 1 ⫺ exp(⫺107.9x2 /x10.41)

where ƒ(x) is the drying rate as a function of the air mass flow rate x1 and the bed thickness x2. A feasible initial point is (800, 0.5). Suggest a reformulation of the problem that will reduce possible numerical difficulties and enhance efficiency of solution using gradient-based methods. 13.6. The following model has been proposed for the optimization of an alkylation process [Sauer, R. N., et al., Hydro. Proc. Petrol. Refiner., 43, 84 (1964)]: Minimize

ƒ(x) ⫽ 0.063x4x7 ⫺ 5.04x1 ⫺ 0.035x2 ⫺ 10x3 ⫺ 3.36x5

Subject to h1(x) ⫽

x4 ⫺ 112 ⫺ 13.167x8 ⫺ 0.6667x82 ⫽ 0 100x1

h2(x) ⫽ x8 ⫺

x5 ⫹ x2 ⫽0 x1

h3(x) ⫽ x4 ⫺ x1 ⫺ x5 ⫹ 0.22x4 ⫽ 0

冉

h4(x) ⫽ 103x3 ⫺ x4x9

冊

x6 ⫽0 98 ⫺ x6

h5(x) ⫽ x9 ⫺ 35.82 ⫺ 0.222x10 ⫽ 0 h6(x) ⫽ x10 ⫹ 133 ⫺ 3x7 ⫽ 0 h7(x) ⫽ x7 ⫺ 86.35 ⫺ 1.098x8 ⫹ 0.038x28 ⫹

and the bounds

冉冊

1.3 (x6 ⫺ 89) ⫽ 0 4

PROBLEMS

0 ⭐ x1 ⭐ 2000

0 ⭐ x2 ⭐ 16,000

0 ⭐ x5 ⭐ 2000 3 ⭐ x8 ⭐ 12

0 ⭐ x3 ⭐120

85 ⭐ x6 ⭐ 93 1.2 ⭐ x9 ⭐ 4

599

0 ⭐ x4 ⭐ 5000

90 ⭐ x7 ⭐ 95 145 ⭐ x10 ⭐ 162

(a) Generate a starting feasible solution to the problem. (b) Obtain a reformulation of the problem suitable for the application of direct-search methods. (c) Obtain a reformulation of the problem suitable for the application of LP-based methodology. 13.7. In the model defined in problem 13.6, constraints h1(x), h4(x), h5(x), and h7(x) are derived from correlations of experimental data, which naturally have some scatter. (a) Reformulate these quality constraints to reflect the possible error in them. (b) Use a suitable gradient-based method to re-solve the problem for each of the possible relative error levels of 0, 1, 5, and 10 percent in the correlations. (c) Explain the results obtained and compare them to the predictions obtained from the multiplier values at the 0 percent error solution. 13.8. The cost minimization problem for a multimedia filter in a municipal water treatment plant can be posed as follows: Minimize

ƒ(x) ⫽ Q⫺1(0.00656x1 ⫹ 0.002189x2 ⫹ 0.03284x3 ⫹ 2345t⫺1)

Subject to h1(x) ⫽ Q⫺0.34de⫺0.51t⫺1.70L0.53 ⫺ 0.007 ⫽ 0 h2(x) ⫽ Q1.4L0.8d⫺1.1 H⫺1t ⫺ 403.23 ⫽ 0 e h3(x) ⫽ 1.337x1L⫺1 ⫹ 0.5086x2L⫺1 ⫹ 0.4x3L⫺1 ⫺ de ⫽ 0 h4(x) ⫽ H ⫺ Ht ⫹ 1.33 ⫽ 0 h5(x) ⫽ L ⫺ x1 ⫺ x2 ⫺ x3 ⫽ 0 g1(x) ⫽ 45 ⫺ 1.41x1 ⫺ 1.35x2 ⫺ 1.1x3 ⭓ 0 g2(x) ⫽ L ⫺ d3e HtQ ⭓ 0 and the bounds

600


t ⭓ 40

Ht ⭐ 8

Q ⭓ 3.0

The variables are defined as L Q t xi de H Ht

⫽ ⫽ ⫽ ⫽ ⫽ ⫽ ⫽

depth of filter bed filtration rate per unit area of filter filtration time depth of filter medium i(i ⫽ 1, 2, 3) equivalent grain size of the multimedia filter headloss change over filtration period headloss at time t

and all must naturally be nonnegative. (a) Generate a starting feasible solution to the problem. On physical grounds it can be expected that L ⭐ 50, Q ⭐ 10, and t ⭐ 500. (b) Suggest transformations and reformulations of the problem that would make it easier and more stable for solution. 13.9. Compare solution of problem 13.8 using a linearization-based code and a direct-search method (ƒ* ⫽ 18.339). Use the results of the former to conduct a sensitivity analysis at the optimum solution. 13.10. Consider the problem [Hesse, R., Oper. Res., 21, 1267 (1973)] Maximize

ƒ(x) ⫽ 25(x1 ⫺ 2)2 ⫹ (x2 ⫺ 2)2 ⫹ (x3 ⫺ 1)2 ⫹ (x4 ⫺ 1)2 ⫹(x5 ⫺ 1)2 ⫹ (x6 ⫺ 4)2

Subject to g1(x) ⫽ (x3 ⫺ 3)2 ⫹ x4 ⫺ 4 ⭓ 0 g2(x) ⫽ (x5 ⫺ 3)2 ⫹ x6 ⫺ 6 ⭓ 0 g3(x) ⫽ 2 ⫹ x1 ⫺ x2 ⭓ 0 g4(x) ⫽ 2 ⫺ x1 ⫹ 3x2 ⭓ 0 g5(x) ⫽ 6 ⫺ x1 ⫺ x2 ⭓ 0 g6(x) ⫽ x1 ⫹ x2 ⫺ 2 ⭓ 0 and the bounds 1 ⭐ x3 ⭐ 5

x4 ⭐ 6

1 ⭐ x5 ⭐ 5

x6 ⭐ 10

and all variables nonnegative.

PROBLEMS

601

(a) Show that the problem can have multiple local maxima (it has 18). (b) Suggest how you might prove that the point (5, 1, 5, 0, 5, 10) is the global maximum point. 13.11. Use a constrained optimization program of your choice to find at least three local maxima to problem 13.10. Prove that the points satisfy the necessary conditions for a local maximum. 13.12. A local minimum of the problem [Rosen, J. B., and S. Suzuki, Commun. ACM, 8, 113 (1965)] Minimize

ƒ(x) ⫽ x21 ⫹ x22 ⫹ 2x23 ⫹ x24 ⫺ 5x1 ⫺ 5x2 ⫺ 21x3 ⫹ 7x4

Subject to g1(x) ⫽ 8 ⫺ x12 ⫺ x22 ⫺ x23 ⫺ x24 ⫺ x1 ⫹ x2 ⫺ x3 ⫹ x4 ⭓ 0 g2(x) ⫽ 10 ⫺ x21 ⫺ 2x22 ⫺ x32 ⫺ 2x42 ⫹ x1 ⫹ x4 ⭓ 0 g3(x) ⫽ 5 ⫺ 2x21 ⫺ x22 ⫺ x32 ⫺ 2x1 ⫹ x2 ⫹ x4 ⭓ 0 is x* ⫽ (0, 1, 2, ⫺ 1) with ƒ* ⫽ ⫺44. (a) Is this the global minimum? Are other local minima possible? (b) Perform a sensitivity analysis to estimate the changes in ƒ* with 5 percent changes in the constraint constants 8, 10, and 5. 13.13. Consider the problem Maximize

ƒ(x) ⫽ x21 ⫹ x22 ⫹ x32 ⫹ (x3 ⫺ x4)2

Subject to g1(x) ⫽ 16 ⫺ 4(x1 ⫺ –12 )2 ⫺ 2(x2 ⫺ 0.2)2 ⫺x32 ⫺ 0.1x1x2 ⫺ 0.2x2 x3 ⭓ 0

g2(x) ⫽ 2x12 ⫹ x22 ⫺ 2x32 ⫺ 2 ⭓ 0 g3(x) ⫽ x2 ⫹ x4 ⭓ 0 (a) Check what bounds may be appropriate in the problem formulation. (b) Check whether the problem is likely to possess multiple local solutions. (c) Generate a feasible starting point. 13.14. Solve problem 13.13 with the bounds ⫺2.3 ⭐ xi ⭐ 2.7

i ⫽ 1, . . . , 4

using a suitable constrained optimization code (ƒ* ⫺ 11.68; Luus, R., AICHE J., 20, 608 (1974)].

602


13.15. Consider the problem ⫺1 ƒ(x) ⫽ x1x23 ⫹ x⫺1 2 x3

Minimize

Subject to g(x) ⫽ x1 ⫺ x32 x3 ⫺ x1x2 x3 ⭓ 0 all xi ⭓ 0 (a) Show that the problem has a nonunique solution. (Hint: Let z1 ⫽ x1x23 and z2 ⫽ x2 x3.) (b) Solve the resulting problem algebraically to show that a family of values of the xi will yield the same optimal objective function value. 13.16. Show that the problem Minimize

ƒ(x) ⫽ 2x1x2 ⫹ x1⫺2x2⫺2

Subject to g(x) ⫽ 1 ⫺ –12 x1 ⫺ –13 x2 ⭓ 0 x1, x2 ⭓ 0 attains its optimum value of ƒ(x*) ⫽ 3 at a family of points satisfying x1x2 ⫽ 1 and g(x) ⭓ 0. 13.17. Suggest a reformulation of the problem Minimize

ƒ(x) ⫽ x21x44

Subject to h(x) ⫽ 0.25 ⫺ x31 / 4x2 ⫽ 0 g(x) ⫽ x1x42 ⫹ x11 / 2x22 ⫺ 1 ⭓ 0 all xi ⭓ 0 that would be easier to solve. 13.18. Generate an initial feasible solution to the problem Minimize

ƒ(x) ⫽ x1x2 x3x4

Subject to g1(x) ⫽ 4x11 / 2 ⫹ x22 x3 ⫺ 1 ⭓ 0 ⫺1 g2(x) ⫽ 2 ⫺ –18 x⫺2 3 ⫺ x1x2 x4 ⭓ 0

g3(x) ⫽ 4 ⫺ 4x21x22 ⫺ x23x24 ⭓ 0 16 ⭓ xi ⭓ 0 Use a random-search method to generate 1000 trial points and record the ratio of feasible to infeasible points generated.

14 ENGINEERING CASE STUDIES In this chapter we present three case studies involving the optimization of large-scale engineering systems. The first study concerns the optimal location of coal-blending plants within a region composed of a number of coal mines and coal users. The second involves the optimal preliminary design of a chemical plant for the production of ethylene oxide and ethylene glycol. The third considers the design of a compressed air storage system to balance peak demands on electric power plants. Each case study illustrates specific formulation and model preparation strategies appropriate to that system. Because of space limitations, the model formulation elements are given only in outline; however, the reader is directed to references that contain the full model details. In addition to discussing the selection of optimization methods, the case studies exemplify various stratagems for enhancing solution efficiency, for treating discrete or 0–1 variables, and for reducing the effective dimensionality of the optimization problem. This chapter thus demonstrates the application of the techniques discussed in Chapter 13 and the algorithms developed in Chapters 1–12 to three real engineering examples. The chapter appropriately serves as a capstone to our study of engineering optimization.

14.1 OPTIMAL LOCATION OF COAL-BLENDING PLANTS BY MIXED-INTEGER PROGRAMMING

In this case study we describe an application of MIP for solving a facility location problem. The study conducted by Ravindran and Hanline [1] examined the feasibility of establishing centralized coal-blending plants in Indiana that could take coals of varying sulfur content and produce a coal Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

603

604

ENGINEERING CASE STUDIES

product that would meet the state’s coal consumer while meeting SO2 emission standards. The case study will also illustrate the use of binary variables for handling fixed costs in the objective function. 14.1.1

Problem Description

Coal is one of the keys to making the United States independent of the uncertain supply of foreign oil. However, a major environmental problem in relying on coal is the presence of sulfur in coal as its is mined. Most midwestern and eastern coal reserves have high sulfur content [2]. Coal-fired furnaces are the major source of sulfur dioxide (SO2) pollution, which is harmful to crops and animals. For this reason, strict SO2 emission regulations have been set for each state by the U.S. Environmental Protection Agency (EPA) [3]. Currently coal users located in areas governed by the most restrictive SO2 standards have limited options such as utilizing flue gas desulfurization techniques (scrubbers) or burning low-sulfur coal. According to the National Rural Electric Cooperative Association [4], ‘‘stack gas scrubbers,’’ which the utilities are forced to install, produced 8–9 ft3 of mudlike sludge for every ton of coal burned, sludge that must be disposed of, creating another environmental problem. Although the option to use low-sulfur coal has been attractive to some utilities and industries, the increased utilization of low-sulfur coal has resulted in higher coal prices for western coal and depressed market for midwestern high-sulfur coal. The option studied here is the establishment of centralized coal-blending plants that could take coals of varying sulfur content (western, midwestern, and eastern coals) and produce a coal blend that would meet customers’ needs while complying with environmental standards. 14.1.2

Model Formulation

Low-sulfur coal and high-sulfur coal are mined at different locations. By a mechanical crushing process these coals are blended to form environmentally acceptable mixtures for consumers. The questions to be studied are (1) How many blending plants should be established? (2) Where should these plants be located? (3) Which consumers and which mines should each blending plant expect to serve? (4) What are the optimal blending ratios of different coals that will meet EPA standards? A single-period (static) model is developed to answer these questions. The combinatorial nature of the site selection problem introduces 0–1 binary variables, and the final formulation reduces to a mixedinteger LP model. Decision Variables. Indices i, j, and k are used as subscripts to denote mine sources, blending plant sites, and consumer locations, respectively. Lowercase letters will be used to denote variables and capital letters for input parameters and constants. The decision variables are defined as follows:

14.1

OPTIMAL LOCATION OF COAL-BLENDING PLANTS

605

vij ⫽ volume (tons) shipped from source i to blending plant j ujk ⫽ volume shipped from blending plant site j to consumer k wij ⫽ (0, 1) binary variable. If coal is shipped from source i to site j, wij ⫽ 1. Otherwise, wij ⫽ 0. xjk ⫽ (0, 1) binary variable. If coal is shipped from site j to consumer k, xjk ⫽ 1. Otherwise, xjk ⫽ 0. yj ⫽ (0, 1) binary variable. If site j is used, yj ⫽ 1. Otherwise yj ⫽ 0. System Objective. The objective is to minimize the total system costs, which include the following:

1. Costs of different coals at the mine source in dollars per ton. 2. Blending costs, which consist of a fixed cost of constructing the blending plant and a variable cost proportional to the volume of coal blended. A blending facility requires a large coal storage area and materialhandling equipment such as a conveyor system, weighing and analyzing equipment, and mixing units. The construction cost, along with other fixed costs, is discounted over the expected life of the facility on an annual basis. 3. Transportation costs, which are a function of origin, destination, and volume shipped. The pricing mechanism for rail-freight rates are very complex, with several different rates, depending on the volume and on whether railroad cars are railroad owned, privately owned, or leased. Often, special rate structures are also used (e.g., unit train rates). In this study, we assume that the transportation costs consist of a fixed cost independent of the volume shipped and a variable cost proportional to the distance and volume shipped. The fixed transportation cost also allows for the expansion of the current transportation system or building new transportation modes. For example, fixed cost may include (a) acquisition of new rail cars instead of using those owned by railroads (the freight rates are much lower for consumer-owned cars), (b) acquisition of new trucks, and (c) building of slurry pipelines for coal transport. Fixed costs are converted to an annual basis by discounting over the expected life of the invested equipment. Thus, the system objective becomes: Minimize

ƒ(v, u, w, x, y) ⫽

冘D 冉冘v 冊 ⫹ 冘C y ⫹ 冘B 冉冘v 冊 ⫹ 冘冘(S w ⫹ 冘冘(F x ⫹ G u ) i

ij

i

j j

j

j

j

j

j

k

ij

ij

i

i

jk jk

jk jk

ij

⫹ Tijvij)

j

(14.1)

606


where Di Sij Tij Cj Bj Fjk Gjk

⫽ ⫽ ⫽ ⫽ ⫽ ⫽ ⫽

mine-mouth cost of coal at source i per unit volume, tons fixed cost of shipping from source i to blending plant j cost of shipping a unit volume from source i to plant j annual fixed cost of blending in plant j variable cost of blending in plant j per unit volume fixed cost of shipping from blending plant j to consumer k cost of shipping a unit volume from blending plant j to consumer k

System Requirements. The following constraints describe the modeled sys-

tem: 1. The volume of coal shipped from a mine source cannot exceed its available supply:

冘v

ij

⭐ Mi

for all i

(14.2)

j

where Mi is the supply available at source i. 2. The volume shipped to a consumer should meet the needs (demand) of that consumer:

冘u

jk

⭓ Nk

for all k

(14.3)

j

Notice that this model does not allow for shortages. Coal rationing, or as an alternative coal shortages, can be handled as follows: N⬘k ⭐

冘u

⭐ Nk

jk

j

where N⬘k is the minimum acceptable demand and Nk the full demand. In this case, a shortage cost would become part of the cost function. Assuming a linear shortage cost, the additional shortage cost term would be

冉

Ak Nk ⫺

冘u 冊 jk

j

where Ak is the shortage cost for consumer k per unit volume. 3. The volume shipped from a blending plant cannot exceed the capacity of that blending plant:

冘u k

jk

⭐ Pjyj

for all j

(14.4)

14.1


607

where Pj is the capacity of blending plant j (assume zero inventory of blended coal) and yj ⫽ 1 if plant j is operational. 4. The total volume of usable coal shipped from a blending plant does not exceed the volume shipped from the mines to the blending plant times a conversion factor:

冘

E vij ⭓ i

冘u

jk

for all j

(14.5)

k

where E is the efficiency of the blending process. 5. It is desirable that constraints assure that the sulfur content of the blended coal shipped to a consumer from a blending plant meets the consumer’s specifications. For this study, we have assumed that the different coals used in the blend combine in a linear relationship. This assumption is made, since empirical data are unavailable to prove or disprove it. If Hi is the sulfur content (percent sulfur by weight) of coal from source i, then the net sulfur content of the blended coal at the blending plant j is 兺iHivij / 兺ivij. If the blending plant j supplies customer k (i.e., if xjk ⫽ 1), then we require 兺iHivij / 兺ivij ⭐ Lk, where Lk is the maximum sulfur content of coal acceptable to customer k. This conditional constraint can be written as follows:

冘H v

i ij

i

冘v

⫺ Lk

ij

⭐ M(1 ⫺ xjk)

for all j and k

(14.6)

i

where M is a large positive number. This constraint forces the blending plant to make a mixture that will meet the most stringent requirements of all the customers it serves. Note that when xjk ⫽ 1, inequality (14.6) reduces to 兺iHivij / 兺vij ⭐ Lk, that is, the required sulfur content of customer k will be met by the net sulfur content of blended coal at site j. On the other hand, if customer k is not supplied from site j, then inequality (14.6) becomes inactive, with xjk ⫽ 0. 6. Miscellaneous constraints: (a) The following constraints ensure that fixed transportation costs are included in the model: ujk ⫺ Mxjk ⭐ 0

for all j and k

(14.7)

vij ⫺ Mwij ⭐ 0

for all i and j

(14.8)

where M is a large positive number. (b) wij, xjk, yj are 0–1 integer variables. (c) vij, ujk are nonnegative.

608


Data. There are 17 coal mines in Indiana that account for 94 percent of the

coal produced in the state [2]. Coal data were collected on all known utilities, industries, and state institutions (e.g., prisons, hospitals, universities, schools) that use over 1000 tons of coal annually [2, 5]. This included 33 utility plants and 46 industrial-institutional customers. However, because of computer size limitations, Indiana was divided into 15 coal-consuming regions. These regions were chosen based upon consumer demand patterns. Both the coke and the cement industries have special coal requirements and hence were not included in this study. Twelve possible blending plant sites were chosen based on discussions with the Indiana state energy officials [5]. Figure 14.1 illustrates the location of coal mines, major consumer locations, and blending plant sites used in the study. The existing transportation system was assumed to be adequate for coal movement in Indiana, and fixed transportation costs were not included. The variable transportation costs were based upon average coal freight rates (cents per ton-mile) published in the Coal Industry Manual [2]. Using the coal freight rates and the distance between the origin and destination, the trans-

Figure 14.1 Indiana map showing regions, source and sink sites, and blend plant locations.

14.1


609

portation cost coefficients were computed [1]. Railroad mileages were estimated from railroad data available [6]. To calculate the sulfur content of coal from various mines, the mean sulfur content of coal seams was found for each mine based on the Geological Survey Report of the State of Indiana [7]. An estimate was made of the acceptable sulfur level in the coal burned by each consumer. The method of estimation is based on the EPA standards as they relate to sulfur dioxide emissions and is described in reference 1. As there were no blending plants commercially available, restrictions pertaining to blending plant capacities were ignored [Eq. (14.4)]. It was also assumed that any loss of coal due to the blending process was negligible. Hence, there was a conservation of mass through each blending plant, and E was set to 1.0 in Eq. (14.5). Coal-blending plants were estimated to cost up to $12 million. The model was solved over a cost of $1.8–11.3 million. Life expectancy of a coal plant was assumed to be 30 years. The cost was discounted over 30 years at a discounting factor of 10 percent, thus giving the annual cost Cj. The operating costs Bj (excluding coal costs) of blending plants were estimated at $0.25 per ton of coal blended. (These estimates were made by a large utility company that had contemplated building a blending facility.) 14.1.3

Results

An illustrative example is shown in which the annual fixed cost of a blending facility is $130,000. With an expected lifetime of 30 years and a discounting factor of 10 percent, this is equivalent to an investment of approximately $11.3 million. The final mixed-integer program had 60 constraints, 406 variables, 12 binary variables, and 1767 nonzero elements. A computer code called MIPZI [8] based on Balas’s implicit enumeration method [9] was used to solve the problem. It took 56.8 sec of CPU time on the CDC 6500 computer at Purdue University. The results show that six blending plant sites were selected for use. The six plant sites and their capacities are as follows:

Site

Location

Volume Handled, Tons Coal

Consumer Areas Served

1 2 6 8 10 11

Gary South Bend Logansport Indianapolis Terre Haute Evansville

4,895,000 772,900 3,529,000 6,792,000 9,729,100 5,610,900

1 2 6 3, 7 4, 8, 10, 14 5, 9, 11, 12, 13, 15

An analysis of these results shows that all the coal produced in Indiana is used and an additional 9 million tons is brought in. Blending plant 2 gets

610


minimum volume, yet it is still economical to build the plant. The EPA requirements are met or exceeded at all blending sites. Since there were no commercially operated blending plants in existence, a sensitivity analysis with respect to the blending plant costs was done. The model was run with lower fixed costs for blending plants; however, the optimum number of sitings did not change. This is attributable to the huge volume of coal shipped. Blending plant costs were not very significant in comparison to transportation costs. When the model was altered so that all 12 blending plant sites were to be used to supply the coal to the surrounding consumers, the annual costs increased from $45.1 million to $51.7 million, a 14.5 percent increase. This increase consists of $0.8 million for blending plant costs and $5.8 million for additional transportation costs. This emphasizes the importance of careful considerations and selection of blending plant sites.

14.2 OPTIMIZATION OF AN ETHYLENE GLYCOL–ETHYLENE OXIDE PROCESS

In this section we outline the development and optimization of a simplified model of a chemical process. Space limitations preclude the complete elaboration of the process model and the presentation of constraint sequencing and reformulation mechanics. These details are given in references 10 and 11, including a listing of the final Fortran subroutines used in the optimization runs. The objective of the present discussion is to outline the form of the model and to highlight the tasks involved in arriving at a successful solution of the optimization problem.

14.2.1

Problem Description

A preliminary design optimization is to be performed for a plant to produce ethylene oxide by direct oxidation of ethylene with oxygen and coproduce ethylene glycol by hydration of a portion of the ethylene oxide. The key plant design parameters are to be selected so as to maximize net profit before taxes. In the process, shown schematically in Figure 14.2, ethylene, oxygen, nitrogen, and recycle gasses are mixed and fed to a tubular reactor packed with silver catalyst. Nitrogen is used as a diluent to keep the mixture outside the flammability limits. The oxidation reactions C2H4 ⫹ –21 O2 → C2H4O and

611

Figure 14.2 Process for the manufacture of ethylene oxide and ethylene glycol.

612


C2H4 ⫹ 3O2 → 2CO2 ⫹ 2H2O are highly exothermic, and temperature is controlled by transferring heat to boiling Dowtherm. The reactor effluent gases are fed to an absorber in which water is used to absorb C2H4O. Except for a small vent, the residual gases are sent to a carbon dioxide removal system and are then compressed to the operating pressure of the reactor. The solution from the ethylene oxide absorber is fed to a stripping column, where steam is used to remove the ethylene oxide. Most of the bottom stream from the stripper is returned to the top of the absorber and the remainder is sent to the glycol reactor. The ethylene oxide–rich stream from the top of the stripper is compressed and condensed and sent to the light-ends removal column in which residual carbon dioxide and other minor undesirable constituents are removed. The bottom from this column now contains only ethylene oxide and water. A fraction of this stream is sent to the final purification column, where pure ethylene oxide is recovered as overhead product. The streams from the stripper and the light-ends column are mixed with makeup water and fed to a stirred tank reactor. Here ethylene oxide is hydrated to form mono, di, tri, and higher polyglycols in the presence of sulfuric acid. Monoglycol is the preferred product. The reactor effluent is passed through a sulfuric acid neutralization system and then fed to an evaporator. The bulk of the water is removed in the evaporator, and the remainder in a distillation column. The mixture is then fed to a final purification column, where monoglycol is recovered as overhead product. 14.2.2

Model Formulation

The model for this system consists basically of a set of equations and relations describing the individual processing units. In general, each such set will involve the following categories of equations: material balances, energy balances, physical property correlations, equipment design equations, and various types of inequality constraints. In this section we will only present selected examples of these types of equations. A more complete model formulation is given in references 10 and 11. Material Balance Equations. The simplest of the model equations will usually

be the material balance equations. These are typically linear in the species flows and, in general, one balance equation can be written for each species processed in a unit. For example, in the case of the absorber, the balances on CO2 and on ethylene oxide take the form CO6 ⫹ CO9 ⫽ CO5 EO6 ⫹ EO9 ⫽ EO5 ⫹ EO8

14.2

613

OPTIMIZATION OF AN ETHYLENE GLYCOL–ETHYLENE OXIDE PROCESS

where COi ⫽ flow of CO2 in stream i, mol/hr EOi ⫽ flow of C2H4O in stream i, mol/hr Energy Balance Equations. The energy balance equations in chemical process applications typically exclude mechanical energy flows and therefore consider only enthalpy changes. In cases in which thermal properties such as heat capacities or latent heats can be assumed to be constant with temperature, the balances reduce to, at worst, quadratic equations in the temperatures and species flows. In other cases, constant properties may not be assumed, and then the balance equation will include additional nonlinear terms. For instance, in the ethylene oxide reactor, the isothermal heat load Q will be given by

冘 (N5 ⫺ N4) 冕 C T

Q ⫽ ⌬H0 ⫹

i

all i

77⬚

Pi

dt

where the first term corresponds to the enthalpy change due to reaction and the second term corresponds to the difference in the sensible heat of the products and reactants. In particular, ⌬H0 ⫽ (EO5 ⫺ EO4)45,300 ⫹ (CO5 ⫺ CO4)284,000

Each CPi is the gas heat capacity of species i given by CPi ⫽ ␣i ⫹ ␤iT ⫹ ␥iT 2 where ␣i, ␤i, and ␥i are empirical constants and N5i, N4i are the molar flows of species i in streams 5 and 4, respectively. Equilibrium Relations. In addition to the thermal properties, chemical process models also require vapor–liquid equilibrium distribution coefficients. These can be rather complex empirical expressions such as the following expression for the Henry’s law constant for CO2, kCO2 [12]:

冉

ln(kCO2) ⫽ 16.43 ⫺ ⫹

冉冊冉 P R

2

冊冉

冊冊

2698 P 203.14 9.7.56 ⫻ 104 ⫺ ⫺ ⫺ 0.00662R T R T T2

119.76 552.22 3.964 ⫻ 107 9.518 ⫻ 108 ⫹ ⫺ ⫹ T T2 T3 T4

where T, P, and R are the temperature, pressure, and universal gas constant, respectively.

614


Equipment Design Equations. The equipment sizing or design equations range from algebraic equations to complex sets of differential equations. The former either can be expressions in which design variables can be explicitly evaluated given the other unit parameters or else involve implicit expressions that would require iterative solution for their evaluation. For instance, the design of the ethylene oxide absorber can be modeled in a straightforward manner, assuming constant pressure and isothermal operation, and using the Kremser equation [13] to calculate the number of theoretical stages N. With ethylene oxide as the key species, this equation takes the form

冋冉

ln N⫽

冊冉冊册

YEO6 ⫺ mXEO8 YEO5 ⫺ mXEO8

1⫺

1 1 ⫹ A A

ln A

(14.9)

where m is the equilibrium constant for ethylene oxide in water (a function of temperature and pressure); A ⫽ L/mG, where L/G is the ratio of liquid to gas molar flows in the absorber; and YEOi and XEOi are the vapor and liquid mole fractions of C2H4O in stream i. By contrast, the design equations for the water removal column, although also algebraic, are more complex. The design equations for preliminary design of multicomponent distillation columns consist of the Fenske equations, the Underwood equations, and the Gilliland correlation [14]. In the present case, the column can be designed with water and monoglycol as the key components. Using the Fenske equation, the minimum number of theoretical separation stages, Nm, is calculated as follows: Nm ⫽ where ␣w1 ␣tw1 ␣bw1 Tb Tt pi(T)

log[(W18/EG118) ⫻ (EG119/W19)] log ␣w

(14.10)

(␣tw1 ⫻ ␣bw1)1 / 2 pw(Tt)/pEG1(Tt) pw(Tb)/pEG1(Tb) temperature at bottom of column temperature at top of column partial pressure of species i expressed as an empirical function of temperature T EGlj, Wj ⫽ monoglycol and water flows in stream j ⫽ ⫽ ⫽ ⫽ ⫽ ⫽

The temperatures at the top and bottom of the column are related to the compositions at the corresponding locations by assuming that stream 18 is at its dewpoint and stream 19 at its bubble point temperature. This gives rise to two equations with implicit temperature dependence,

14.2


615

冘

1 p (T )Xi17 ⫽ 1.0 P i i b ⫽ 1.0 冘 pYi18 (T )

P

i

i

(14.11)

t

where i ⫽ water, monoglycol, diglycol P ⫽ total column pressure The minimum reflux ratio is calculated using the Underwood equation. Assuming saturated liquid feed, the equation ⫽0 冘 ␣␣ Xi17 ⫺␪

with ␣w1 ⭐ ␪ ⭐ ␣21

i

i

(14.12)

i

fixes the value of the Underwood parameter ␪. The equation Rm ⫹ 1 ⫽

冘 ␣␣ Xi18 ⫺␪ i1

(14.13)

i1

serves to define the minimum reflux ratio Rm. The actual reflux is usually selected as a design multiple of Rm or can itself be selected by optimization. With R known, the total condenser cooling duty and reboiler heating duty can be calculated from simple energy balances. Finally, with Nm, Rm, and R determined, the actual number of stages can be established using Eduljee’s analytical expression [15] for the Gilliland correlation:

冋冉

冊册

N ⫺ Nm R ⫺ Rm ⫽ 0.75 1 ⫺ N⫹1 R⫹1

0.5668

(14.14)

The design equations for the tubular ethylene reactor with external wall cooling will consist of a system of differential equations describing the changes in the species flows, stream enthalpy, and stream pressure as the feed mixture flows through the catalyst tubes. For instance, the production of C2H4O will be given by the equation dNEO ⫽ rEO dw where w is the weight of catalyst; rEO is the rate equation for the ethylene oxide reaction given by reference 16,

616


rEO ⫽ F⫽

k1pE pO1 /22 䡠 F lb 䡠 mol of EO 1 ⫹ 24.55pE (hr)(lb of catalyst) 1 1 ⫹ 30.63pH2O ⫹ 7.68(pCO2 ⫹ pEO)

and k1 is the temperature-dependent rate parameter

冋冉

k1 ⫽ 0.09428 exp 13.41 1 ⫺

冊册

923 T

In these expressions, pE, pO2, pH2O, pEO are partial pressures of ethylene, oxygen, water, and ethylene oxide, respectively. Similar rate expressions are used for the production of CO2. The system of differential equations is not analytically integratable. As an approximation that will allow the reactor model to be more readily treated by NLP methodology, the numerical solutions of the differential equations generated over a range of anticipated operating conditions were fit with response surfaces of the form w ⫽ A0YE4A1YO4A2XA3PA4(e⫺T ) A5 ln(2 ⫺ X) A6

(14.15)

for the catalyst weight w and an equation of the form S⫽

1 1 ⫹ A7YO4A8

for the selectivity S. The parameters A0 through A8 are regression constants, X is the overall conversion of ethylene in the reactor, and YE4 in the mole fraction of ethylene and YO4 the mole fraction of oxygen in stream 4. The response surfaces inherently have only a limited range of validity and hence may need to be refitted if their limits are exceeded during use of the model. Inequality Constraints. In addition to the large number of nonlinear and linear

equations of the type described above, the process model also has associated with it a number of inequalities. These can be grouped into three classes: 1. The explicit constraints that are an integral part of the model 2. The implicit constraints required to avoid numerical difficulties 3. Upper and lower bounds on key variables that are imposed to restrict the model to a given domain for a particular application. The most important explicit constraints are the flammability concentration limits on the ethylene oxide reactor feed, the bounds on the Underwood parameter ␪ in Eq. (14.12), and the requirement that the saturated temperature

14.2


617

of the stripping steam be above the stripper feed temperature to avoid excessive steam condensation. The implicit constraints consist of the arguments of various logarithmic and power functions. In principle, all such arguments should be bounded to avoid possible arithmetic errors or undefined function values. However, in practice the majority of the resulting large number of inequality constraints can be deleted during initial model runs, since the argument values may be very far from the critical values. In the present case the following were found to be important: 1. The arguments of the logarithmic terms in the Kremser equation for the EO stripper and absorber [Eq. (14.9)] 2. The argument of the Gilliland correlation in the case of the glycol purification column [Eq. (14.14)] 3. The minimum reflux for the glycol–water column 4. The temperature differences in the absorber liquid heat exchanger and in the evaporator The two temperature difference constraints are necessary to avoid calculation of negative heat transfer rates and, thus, negative heat transfer areas. The third type of inequality constraint is the set of upper and lower bounds on various model variables, which will be specific for the given application. Objective Function. For the present study, the objective function consists of the net profit before taxes expressed on an hourly basis. Net profit is sales revenue minus total annual cost. The latter consists of utility and raw material costs plus an appropriate fraction of the total installed equipment cost. The total installed equipment cost is the sum of the costs of each of the major items of equipment. These individual equipment costs are estimated from cost–capacity data for various equipment types reported by Guthrie [17] fitted to a power function. For example, the installed cost of a tower of stainless steel–clad construction with stainless steel bubble-cap trays is given by

2{(674 ⫹ 4.63P)[D1.021H 0.965(D ⫹ H)⫺0.175] ⫹ [88.42D1.454H1.13(D ⫹ H)⫺0.275]}

where P, D, and H are the operating pressure, diameter, and height of the tower, respectively. The utility costs include the costs of steam for stripping, evaporation, and reboiling; electricity for the recycle gas compressor; process water for the glycol reactor; and sulfuric acid. The raw material costs consist of the costs of ethylene, oxygen, and nitrogen. Particular values of these various parameters are given in reference 10.

618

14.2.3


Problem Preparation

The complete equation-oriented flowsheet model consists of more than 100 equations and an even larger number of bounds and inequalities. Because of the network structure of the underlying physical problem, the problem variables and equations can be sequenced for solution so that only a reduced set of 22 independent variables and associated constraints has to be explicitly considered. Two types of equations resisted this type of sequencing: (1) the material and energy balance equations required to close key recycle loops and (2) subsets of the individual equipment design equations that could not be solved explicitly for one or more of the design variables. The former consisted of the material balances associated with the ethylene oxide loop and the energy balance associated with the absorber–stripper loop. The equipment equations that could not be sequenced are those associated with the glycol–water column, the bubble and dewpoint equations, as well as the instances of the Underwood equation (14.12). The resulting NLP problem thus consists of 22 variables, 7 equality constraints, 12 inequality constraints, and a virtually complete set of upper and lower bounds. Note that, by virtue of the nonlinear equality constraints and because of the presence of numerous power-law terms in the objective function, the problem is nonconvex and highly likely to possess multiple local solutions.

14.2.4

Discussion of Optimization Runs

The problem was implemented for optimization using a GRG code [18]. The initial runs with GRG using the original model were unsuccessful, often resulting in negative arguments for power and logarithmic functions and in premature termination because of scaling difficulties. These problems were remedied in part by imposing the class 2 inequality constraints discussed in Section 14.2.2. Three of the variables had to be scaled: the variables corresponding to the nitrogen feed rate (multiplied by l03), the ratio of water to C2H4O in the glycol reactor (multiplied by 102), and the fraction of C2H4O sent to the purification column (multiplied by 102). With these changes, difficulties were still encountered with arithmetic errors and invalid arguments during the constraint adjustment phases of the algorithm. Apparently, the modest constraint infeasibilities permitted by the algorithm during this phase were sufficient to cause abortion of the run. To surmount this type of difficulty, the model was modified by adding checks of function arguments and resetting these arguments to acceptable lower bounds if invalid arguments were encountered. Such resetting, in principle, introduces derivative discontinuities at the reset points; however, since the optimal solution was suffi-

14.2


619

ciently far from the reset points, difficulties with discontinuities were not observed in the runs. The modified model was successfully optimized using the GRG code from four different starting points. As shown in Table 14.1, a typical optimization run required less than 100 sec CPU time on a CDC 6500. In three out of four cases the objective function attained the same value, 2938.43, and in the fourth it differed only slightly. Hence, the objective value of 2938.43 is probably the desired global optimum. Three variables—the fraction of ethylene oxide sent to the final purification column, the operating pressure in the ethylene oxide stripper, and the liquid– gas ratio in the ethylene oxide absorber—reached their lower bound. Four variables—conversion in the ethylene oxide reactor, the temperatures in the two reactors, and the fraction of water removed in the evaporator—reached their upper bounds. Fifteen variables are within their bounds. The constraints on the flammability limits and two temperature differences are binding at the optimum. The temperatures in the reactors and the conversion of ethylene are expected to reach their upper bounds, because higher values of these variables favor higher production of ethylene oxide and ethylene glycol, which in turn increases the profit. The upper bound value for the fraction of water removed in the evaporator suggests that at that upper bound value it is still cheaper to remove water in the evaporator rather than in a distillation column. The flow split, X(6), determines the fraction of ethylene oxide purified and sold as ethylene oxide product, the remainder being converted to ethylene glycol. Its optimum value reached the lower bound of 30 percent, which implies that it is more profitable to sell ethylene oxide as ethylene glycol than as ethylene oxide. This can be expected in view of the higher price of 19¢/lb used for the former compared to l5¢/lb used for ethylene oxide. As described by Sarma and Reklaitis [10], further runs were carried out both to test the validity of the model and to study the effects of parameter changes on the optimal design. Such studies are important to support the credibility of the model as well as to indicate model elements that may require more detailed study using a rigorous process flowsheet simulator [11]. In summary, the lessons to be found in this case study are that model formulation and preparation for solution are difficult and time-consuming steps that require detailed knowledge of the system being studied. Even after careful preparation, it proved necessary to scale some of the variables and to add some inequality constraints to avoid arithmetic errors and invalid arguments. Since the GRG algorithm does allow infeasible excursions, it was necessary to add checks to the model subroutine to prohibit excursions that would cause fatal arithmetic errors. Finally, once the problem would successfully execute, it proved prudent to repeat solution with several different starting points to increase confidence in the solution accepted as optimal.

620

4952.8 1.094 240.0 0.1938 2.4 34.0 0.37 14.1 14.0 44.4 2.1 2.42 0.55 44.4 2410.0 74.0 0.067 0.89 159.0 68.0 1.36 .276 2657.22 —

1

Optimization Results

X(1) X(2) X(3) X(4) X(5) X(6) X(7) X(8) X(9) X(10) X(11) X(12) X(13) X(14) X(15) X(16) X(17) X(18) X(19) X(20) X(21) X(22) Net profit CPU time, sec

Variable

Table 14.1

5892.785 1.552 219.99 0.1613 4.4893 38.997 0.3737 30.684 14.0 19.99 2.8636 1.888 0.5 49.61 2800.0 66.0 0.0575 0.9046 175.798 82.21 0.138 0.5 2100.7 —

2 12901.7 0.75 220.0 0.068 4.5 39.0 0.35 30.0 14.0 20.0 3.0 2.0 0.5 50.0 2800.0 66.0 0.05 0.9 177.0 85.0 0.14 0.5 1505.02 —

3

Starting Points

14237.1 0.75 220.04 0.060 4.66 35.0 0.34 23.62 14.0 23.65 2.99 1.92 0.5 51.06 2744.4 66.0 0.0522 0.9 177.3 83.44 0.1399 0.5263 1653.38 —

4 3257.33 0.6667 242.0 0.3 1.0362 30.0 0.385 5.428 12.72 44.41 2.0 2.4235 0.5 44.41 2753.58 75.0 0.0173 0.98 160.62 70.95 2.518 .2843 2936.23 60.1

1 3257.33 0.6667 242.0 0.3 1.0362 30.0 0.385 5.428 12.72 44.41 2.0 2.4235 0.5 44.41 2756.03 75.0 0.0173 0.98 176.4 83.62 0.1429 0.492 2938.43 63.8

2 3257.33 0.6667 242.0 0.3 1.0362 30.0 0.385 5.428 12.72 44.41 2.0 2.4235 0.5 44.41 2756.03 75.0 0.0173 0.98 176.41 83.63 0.1429 0.492 2938.43 75.8

3

Corresponding Optimal Points

44.41 2756.03 75.0 0.0173 0.98 176.41 86.63 0.1429 0.492 2938.43 93.7

3257.33 0.6667 242.0 0.3 0.0632 30.0 0.385 5.428 12.72 44.41 2.0 2.4235

4

14.3

OPTIMAL DESIGN OF A COMPRESSED AIR ENERGY STORAGE SYSTEM

621

14.3 OPTIMAL DESIGN OF A COMPRESSED AIR ENERGY STORAGE SYSTEM

Finally in this section we use a case study involving the design of a compressed air energy storage (CAES) system to demonstrate formulation technique and several of the methods and concepts discussed in previous chapters. It is appropriate to devote considerable attention to problem formulation technique, since it remains to a large degree an art which requires considerable practice to master. We demonstrate here a decomposition approach, involving the identification of coupling variables, suitable for application to large NLP problems. The work was carried out in two locations, with the support of the U.S. Department of Energy: by F. W. Ahrens and co-workers at Argonne National Laboratory in Chicago, Illinois, and K. M. Ragsdell and students at Purdue in West Lafayette, Indiana. The employed decomposition approach very significantly enhanced the dual design team interaction, since each site had primary responsibility for a single subsystem optimization. Numerical results are given for a CAES plant employing the Media, Illinois, Galesville aquifer formation. Additional modeling details and results are given in a number of related publications [19–21]. 14.3.1

Problem Description

Compressed air energy storage is one of the technologies that is currently available to electric utilities to supply peak power using stored energy previously generated during periods of excess capacity [22]. A goal of this study was the determination of the economic viability of CAES peaking plants employing below-ground naturally occurring aquifers. The use of energy storage systems can be economically advantageous to utilities, since they improve the utilization of high-efficiency base plants and permit a reduction in the amount of premium fuel required to generate peak power. Consider the CAES power plant diagrammed in Figure 14.3, which is based on a split Brayton cycle and comprises four equipment groups: a reversible motor/generator, air compression equipment, an air-storage reservoir (with associated piping), and power extraction equipment. The couplings on the motor/generator allow the plant operators to use either electrical power from the utility grid to compress air or the stored compressed air and some premium fuel to generate power. The CAES system configuration shown in Figure 14.3 is typical, but many variations in equipment arrangement are possible [22]. In terms of its interaction with the other equipment groups, the turbine system can be characterized by its design inlet pressure (pti) and its mass flow rate per unit power output (m ˙ ⬘). The latter depends on the turbine inlet temperatures (premium fuel consumption) and equipment arrangement and design. Details of these relationships are given by Kim and Kartsounes [23]. Because of the requirements for storing large amounts of high-pressure compressed air (e.g., l07 –l08 ft3 at 50 atm for a typical 200-MW plant), it is

622


Figure 14.3 CAES power plant.

known that underground air reservoirs are an economic necessity. The reservoir can be either a cavern (in hard rock or in a salt dome) or a porous rock layer (most commonly an aquifer), such as an edgewater aquifer. The storage reservoir design requirements (capacity, pressure level, piping design, etc.) are interdependent with the selection (performance characteristics and operating conditions) of the above-ground compression and power generation equipment and the desired power output and operating cycle of the CAES plant. In turn, the decision on the power level and duty cycle is impacted by the economic and technical characteristics of the utility grid, by the cost of premium fuel, and so forth. Thus many technical and economic trade-offs must be considered in specifying a CAES plant design. 14.3.2

Model Formulation

In Figure 14.3 we have seen the basic components of a CAES plant. Although the utility grid is not physically part of the CAES plant, the interaction should be considered, since the cost of base-load power and the utility load cycle may have a strong influence on design cost of the CAES facility. Correspondingly, the CAES costs will influence the cost of power sold by the utility. Decomposition Strategy. We find it convenient, if not essential, to decom-

pose the system into three groups or subsystems as seen in Figure 14.3. Subsystem 1 contains the air storage reservoir, air compression train, and main piping and air distribution system. Subsystem 2 contains the power generation train, and subsystem 3 contains the motor/generator and the utility grid. With

14.3


623

the subsystems formed in this way, it is possible to choose coupling and internal variables so that the subsystems can be designed with a degree of independence from the other subsystems. The exact dependence is contained in the coupling variable relationships. For instance, we assume that subsystem 3 (the utility grid) affects the rest of the system through a variable UL, the utility load cycle, as shown in Figure 14.4. This single variable could, of course, represent many variables in the utility load cycle, but this is not pursued here, since our interest is primarily with subsystems 2 and 3. Finally, we suggest that the direct interactions (or coupling) between subsystems 1 and 2 are dependent on only two variables, the inlet pressure to the power ˙ ⬘. As the figure suggests, generation train pti and the specific mass flow rate m there is also the indirect influence of the utility load cycle. We quantify this effect by choosing the utility load cycle. The criterion for optimal design is chosen to be the total normalized cost C of the system (i.e., cost per unit of electricity generated by the CAES power plant). This total cost is the sum of the individual costs, which include fuel cost, maintenance, and charge rate on capital. The costs have to be minimized subject to various performance and technical constraints. The implication for CAES plant design is that for a given utility load cycle an optimization of subsystem 1 would provide the minimum subsystem operating cost C 01 and values for the corresponding subsystem design variables as a function of the ˙ ⬘. Similar optimization for subsystem 2 would coupling variables pti and m yield C 02 (the minimum operating cost of subsystem 2) and its optimum design as a function of the coupling variables only. Finally, the sum of C 01 and C 02 can be minimized to determine the optimum values of the coupling variables, the minimum plant cost C*, and the optimal plant design.

Figure 14.4 Utility load cycle UL.

624


Subsystem 1: Storage. The aquifer (originally water filled) should have the shape of an inverted saucer to prevent migration of the compressed air. The air bubble is formed by displacing the innate water; the compressed air is contained between the airtight caprock and a bottom layer of water. The compressor train included in this subsystem follows the recommendations of United Technologies Research Center [24]. A detailed discussion of the model employed is given by Ahluwalia et al. [25]. Here, we focus only on the conceptual formulation of the optimal design problem. In the optimization of subsystem I, the objective is to determine the combination of internal design variables that minimizes the subsystem operating ˙ ⬘, and UL. The set of cost for given values of the coupling variables pti, m design variables can be classified into two subsets. The first subset includes variables that are restricted to take a limited number of discrete values. Engineering considerations require that the main piping diameter, the type of low-pressure compressor, and the reservoir well-bore diameters be restricted to discrete, commercially available designs. As the number of alternatives is limited, the simple method of incorporating these discrete variables into the optimization via an exhaustive search in all discrete dimensions is suitable. The remaining internal optimization variables of subsystem 1, which are assumed to be continuous, are the four geometric parameters of the reservoir design, NW, H, Aact, and d illustrated in Figure 14.5, and the energy storage process variables tcbi and tcei. The variables tcbi represent the times during the weekly cycle when energy storage processes begin, and tcei are the ending times of these processes. The storage (charging) time variables are shown for a typical cycle in Figure 14.4. The operating cost, to be minimized, can be written as

冘

C1(NW, H, Aact, d, tcbi, tcei) ⫽ K1(UL)CT ⫹ K2(UL)Pci (tcei ⫺ tcbi)

(14.16)

i

In the equation above, K1 and K2 are functions of the coupling variable UL but are treated as constants for the purpose of optimization. Similar notation is used to represent functions of other coupling variables and functions of the three discrete internal variables. The first term in Eq. (14.16) represents the operating cost due to the annual charge rate on capital, CT, of subsystem 1, where CT is the sum of capital costs of the various components: CT (NW, H, Aact, d, tcbi, tcei) ⫽ WC ⫹ LC ⫹ BC ⫹ CC ⫹ K3(piping)

(14.17)

K3 (piping) is the capital cost of the main piping and distribution system, which depends upon the piping design selected. The capital cost of wells is WC(NW, H, Aact) ⫽ NW[KW1 ⫹ KW2{H ⫺ F(Aact)}]

(14.18)

with constants KW1, KW2, and F(Aact), a known function of Aact, determined from reservoir geometry. The term within curly brackets in Eq. (14.18) is the

14.3

625


Figure 14.5 Aquifer reservoir geometry.

depth to which wells must be bored. The second term on the right in Eq. (14.17) is the cost of purchasing the land over the proposed reservoir, LC(d) ⫽ Kl A(d)

(14.19)

where A(d) is the land area over the air reservoir, a known geometric function of d. In this simplified model, the capital cost of initially displacing water from the aquifer, or bubble development, is calculated in terms of energy required to compress the volume of air in the bubble, which is a function of d. Finally, the capital cost, CC, of the compressor train is expressed as CC(Nw, H, Aact, d, tcbi, tcei) ⫽ Kcl1 ⫹ Kcl2Mc ⫹ Kcb1

冋再

˙ c pc ⫺ 1 M

Kcl3

冎册

Kcb3

(14.20)

626


Here, Kcl1, Kcl2, and Kcl3 are parametric constants determined by the choice of ˙ c is the air mass flow rate during the storage proccompressor train design; M esses, chosen to be the same during all storage processes due to compressor performance considerations. The remaining unknown term in Eq. (14.20) is pc, the discharge pressure required of the compressor train. This pressure can be calculated using the pressure drop models given by Sharma et al. [19]. The second term in Eq. (14.16) is the subsystem operating cost incurred due to compressor power consumption, pc, which is given by ˙c Pc(Nw, H, Aact, d, tcbi, tcei) ⫽ [KP1 ⫹ KP2Pc ⫹ KP3Pc2 ⫹ KP4Pc3]M

(14.21)

Engineering intuition, aquifer geology and geometry, and the utility load cycle suggest bounds and functional constraints on the design variables. These have been completely developed and explained by Ahrens et al. [20] and will not be repeated here. In summary, we place bounds on the storage process times as suggested in Figure 14.4 and upper and lower limits on the four physical variables Nw, H, Aact, and d as defined in Figure 14.5. Functional constraints are imposed that require that all storage processes end after they begin, that the wells be placed close enough to ensure full utilization of reservoir volume, that the well bores not physically interfere one with another, that the ‘‘bubble’’ be no larger than the land purchased, that the wells not be drilled to a depth that would cause ‘‘coning,’’ that the required compressor power be less than that which the utility is willing and/or able to supply at any time, and, finally, that the pressure requirement of subsystem 2 be met at all times by subsystem 1. In summary, there are 16 design variables with 32 bounds and 12 constraints, 4 of which are nonlinear. Subsystem 2: Generation. Subsystem 2 of the CAES system is composed of the high- and low-pressure turbines, their combustors, and the recuperator, as indicated in Figure 14.3. The most interesting design trade-offs for this subsystem are (1) larger, more effective recuperator versus greater premium fuel consumption in the combustors, for preheating the air entering the turbines, and (2) advanced, high-inlet-temperature turbines with high cost but high performance versus conventional, lower temperature, lower cost turbines. An additional trade-off, of secondary importance, is the pressure ratio split between the high-pressure turbine and the low-pressure turbine. The performance model for subsystem 2 is based on a thermodynamic analysis of the components. The detailed equations are given by Kim and Kartsounes [23]. It should be mentioned that the model includes the effect that as the turbine inlet temperatures exceed 1600⬚F it is necessary to use an increased function of the compressed air from storage for the turbine blades and other turbine components. For the purpose of calculating subsystem 2 performance, the coupling var˙ ⬘ (the specific turbine system iables pti (the subsystem inlet pressure) and m air flow rate, lbm/kWh), and Pgen (the total power output from the two

14.3


627

turbines) are regarded as inputs. Because of this, it is not possible to independently specify both turbine inlet temperatures T3 and T5 if fixed, state-ofthe-art turbine efficiencies are assumed. In the present mode, T5 (low-pressure turbine inlet temperature) was considered a design variable and T3, along with several intermediate variables, was subsequently determined during the iterative solution of the model equations. The other design variables of subsystem 2 are the recuperator effectiveness ␧ and the low-pressure turbine pressure ratio rp. The variable rp takes on two discrete values, 11 and 16, corresponding to the current practice of turbomachinery manufacturers. In the optimization of subsystem 2, the objective is to find the combination of internal design variables that minimizes the subsystem operating cost for given values of the coupling variables. During a particular optimization process, the coupling variables pti, m⬘, and UL are fixed, so they will be omitted from the functional relationships that follow. The discrete variable rp is also omitted, since an optimization is performed separately for each of its values. The operating cost to be minimized can be written as ˙ ⬘ ⫹ K0m C(T5, ␧) ⫽ K1(UL)Ccap ⫹ KFQ

(14.22)

The first term represents the operating cost due to the annual charge rate on the capital Ccap of subsystem 2, where Ccap is the sum of capital costs: Ccap(T5, ␧) ⫽ CLGT ⫹ CHGT ⫹ CR ⫹ CBAL

(14.23)

According to expression (14.23), the capital cost is the sum of CLGT, the cost of the low-pressure turbine (including the increased expense of cooling air for high operating temperature); CHGT, the cost of the high pressure turbine; CR, the cost of the recuperator; and CBAL, the cost of the balance of the plant. The second term in Eq. (14.22) is the cost of the premium fuel used in the ˙ ⬘ is dependent combustors. Here KF is taken as $2.50/106 Btu. The heat rate Q on T3(T5, ␧), and T5. The final term in Eq. (14.22) is the operating and maintenance cost of the plant. Finally, we place upper and lower bounds on the three design variables ␧, the recuperator effectiveness, and T3 and T5, the turbine inlet temperatures. 14.3.3

Numerical Results

Using the transformation code BIAS [26] for subsystem 1 and the GRG code OPT [27] for subsystem 2, we have sought the plant design that minimizes the normalized operating cost of generation of 600 MW for 10 hr each weekday. We have employed the Media, Illinois, Galesville aquifer as the storage reservoir. Contour maps and material properties for this aquifer and other problem parameters are given by Sharma et al. [19], Katz and Lady [28], and Ahluwalia et al. [25]. The subsystem 1 problem was solved for a number of ˙ ⬘, using BIAS with automatic scaling. combinations of pti, and m

628


Contours of constant minimized operating cost for subsystem 1 are shown in Figure 14.6. A very significant cost variation is evident. Figure 14.6 indicates that small m ˙ ⬘ values (i.e., low air flow rates) are favored. This is due primarily to the higher cost of the air storage reservoir as the quantity of air stored is increased. The optimum subsystem 1 designs corresponding to points in Figure 14.6 were also found to vary widely. Of particular interest is the number of wells required. It was found to vary from a low of 54 in the lower left (low-cost) region to values in the 200–500 range in the upper right region. Finally, it should be noted that the effects of the discrete variables (low-pressure compressor–compression ratio, well bore diameter, and main pipe diameter) have been studied and are reported by Ahrens [29]. Contours of constant minimized operating cost for subsystem 2 are pre˙ ⬘ values. The minimum cost sented in Figure 14.7 for a range of pti, and m contour (22 mills/kWh) corresponds approximately to designs having the minimum allowed (1500⬚F) turbine inlet temperatures T3 and T5. These correspond to conventional designs proposed for CAES plants. The maximumcost contour (24.5 mills/kWh) shown is near to the constraint boundary representing the upper limit (2400⬚F) on turbine inlet temperatures. The results in Figure 14.7 are based on rp ⫽ 16. It was found that use of rp ⫽ 11 yielded similar but slightly higher cost results throughout the region explored. The optimum recuperator effectiveness ␧ was found to vary from 0.52 to 0.77 for the ranges of coupling variables yielding solutions.

Figure 14.6 Subsystem 1 optimal designs.

14.3


629

Figure 14.7 Subsystem 2 optimal designs.

By the nature of the decomposition strategy employed in this study, the optimum CAES plant (the design that minimizes the power generation cost for the specified utility load cycle and aquifer site) may be easily found by superimposing the results from Figures 14.6 and 14.7. The resulting minimized cost contours are shown in Figure 14.8. Interestingly, even though the individual subsystem contours are open, their sum exhibits an overall optimum that is within the coupling variable domain considered. Figure 14.8 demonstrates that the power generation (operating) cost of the optimum CAES plant is slightly under 35.75 mills/kWh and that the optimum values for the ˙ ⬘ ⫽ 8.5 lbm /kWh. coupling variables are, approximately, pti ⫽ 625 psia and m Knowing the optimum coupling variables, we can readily obtain the optimum values of other design variables from the separate subsystem 1 and 2 optimization results. 14.3.4

Discussion

The optimal design approach affords a significant opportunity for cost savings in the construction and operation of compressed-air energy storage systems, as can be seen from the previously given results. On the other hand, the models necessary to adequately represent such a practical physical system can be quite complex. We have given what we feel to be the least complex system model that will produce a meaningful optimal design. Even with our simplified approach, the complete CAES system optimization (including subsystems

630


Figure 14.8 CAES plant optimizaton results.

1 and 2) involves 20 design variables, 4 discrete design parameters, 8 linear constraints, upper and lower bounds on all design variables, and a nonlinear objective function. Furthermore, the model includes functions that require calculation of the modified Bessel functions of the first and second degree and first and second kind and various spline approximations for empirical data. The decomposition proved to be an effective means of attacking the problem, not only because of the savings in solution time but also because the subsystem optimization problems have utility within themselves. That is, these subgroup results provide insight that would be difficult at best to gather in any other way. Finally, the decomposition strategy employed here allows the use of an orderly modular approach of design. That is, we might envision a different storage system (such as a hard rock cavern), which would produce a different subsystem 1 model. We could perform the subsystem 1 optimizations and synthesize the overall system results just as before. 14.4

SUMMARY

Three case studies in engineering optimization were developed in this chapter. The first resulted in a large-scale linear program that required the introduction

REFERENCES

631

of 0–1 variables to accommodate fixed costs in the objective function. Solution necessitated a branch-and-bound strategy with LP subproblems. The second case study led to a large nonlinear program whose size could be considerably reduced by explicit sequencing of equality constraints. Because of the presence of ‘‘hard’’ constraints whose violation would disallow model evaluation, a GRG solution approach proved effective. The third case study involved a large nonlinear program with some discrete variables and a decomposable structure. The decomposition separated the problem into three subproblems: two involving the optimization of each major subsystem for fixed values of selected coupling variables and the third involving an overall optimization with respect to the coupling variables. Optimization with respect to the discrete variables could be carried out by exhaustive enumeration. The case studies illustrated quite clearly that the solution strategy must be carefully selected to exploit the characteristics of the system under study. Identification of these characteristics requires detailed knowledge of the engineering system as well as a mastery of optimization methodology. These case studies thus reaffirm that the optimization of engineering systems is of necessity the domain of engineers informed in optimization methods. We hope that this book has helped the reader to achieve this state of knowledge.

REFERENCES 1. Ravindran, A., and D. L. Hanline, ‘‘Optimal Location of Coal Blending Plants by Mixed Integer Programming,’’ Am. Inst. Eng. Trans., 12(2), 179–185 (1980). 2. Key Stone Coal Industry Manual, McGraw-Hill, New York, 1975. 3. Compilation of Air Pollutant Emission Factors, 2nd ed., U.S. Environmental Protection Agency, Apr. 1973. 4. Rural Electrification, National Rural Electric Cooperative Association, Washington, DC, Mar. 1978. 5. Indiana Energy Office, State Department of Commerce, Indianapolis, IN, private communication. 6. Rand McNally Railroad Atlas of the United States, Rand McNally & Co., Chicago, IL, 1971. 7. Wier, C. E., Coal Resources of Indiana, Indiana Geological Survey, Bloomington, IN, 1973. 8. MIPZI—Documentation of a Zero-One Mixed Integer Programming Code, Department of Agricultural Economics, Purdue University, West Lafayette, IN, Sept. 1973. 9. Balas, E., ‘‘An Additive Algorithm for Solving Linear Programs with Zero-One Variables,’’ Oper. Res., 13, 517–547 (1965). 10. Sarma, P. V. L. N., and G. V. Reklaitis, ‘‘Optimization of a Complex Chemical Process Using an Equation Oriented Model,’’ Math. Prog. Study, 20, 113–160, 1982.

632


11. Sarma, P. V. L. N., ‘‘Strategy of Modeling and Optimization in Chemical Process Design,’’ Ph.D. dissertation, Purdue University, West Lafayette, IN, 1978. 12. Stewart, P. B., and P. Mungal, ‘‘Correlation Equation for Solubility of Carbon Dioxide in Water, Seawater, and Seawater Concentrates, J. Chem. Eng. Data, 2, 170 (1971). 13. Smith, B. D., Design of Equilibrium Stage Processes, McGraw-Hill, New York, 1963. 14. Treybal, R. E., Mass Transfer Operations, McGraw-Hill, New York, 1968. 15. Eduljee, H. E., ‘‘Equations Replace Gilliland Plot,’’ Hydrocarbon Proc., 54(9), 120 (1975). 16. Buntin, R. R., ’’A Kinetic Study of Oxidation of Ethylene on a Silver Catalyst,’’ Ph.D. thesis, Purdue University, West Lafayette, IN, Aug. 1961. 17. Guthrie, K. M., ‘‘Data and Techniques for Preliminary Capital Cost Estimating,’’ in Modern Cost Engineering Techniques (H. Popper, Ed.), McGraw-Hill, New York, 1970. 18. Lasdon, L. S., A. D. Warren, A. Jam, and M. W. Ratner, ‘‘Design and Testing of Generalized Reduced Gradient Code for Nonlinear Programming,’’ ACM Trans. Math. Software, 4, 34–50 (Mar. 1978). 19. Sharma, A., H. H. Chiu, F. W. Abrens, R. K. Ahluwalia, and K. M. Ragsdell, ‘‘Design of Optimum Compressed Air Energy Storage Systems,’’ Energy, 4, 201– 216 (1979). 20. Ahrens, F. W., A. Sharma, and K. M. Ragsdell, ‘‘Computer Aided Optimal Design of Compressed Air Energy Storage Systems,’’ ASME J. Mech. Des., 102, 437– 445 (1980). 21. Ahrens, F. W., A. Sharma, and K. M. Ragsdell, ‘‘Optimal Design of Compressed Air Energy Storage Systems,’’ Proceedings of the Sixth UMR-DNR Conference on Energy, 431–441 (Oct. 1979). 22. Vosburgh, K. G., ‘‘Compressed Air Energy Storage,’’ AJAA J. Energy, 2(2), 106– 112 (1978). 23. Kim, C. S., and G. T. Kartsounes, ‘‘A Parametric Study of Turbine Systems for Compressed Air Energy Storage Plants,’’ Final Report FY1977, Report No. ANL/ ES-64, Argonne National Laboratory, Chicago, IL, 1978. 24. Davison, W. R., and R. D. Lessard, ‘‘Study of Selected Turbomachinery Components for Compressed Air Storage Systems,’’ Report No. ANL / EES-TM-14, United Technologies Research Center for Argonne National Laboratory, 1977. 25. Ahluwalia, R. K., A. Sharma, and F. W. Ahrens, Proc. Compressed Air Energy Storage Technology Symposium, Pacific Grove, CA, 1979, pp. 295–325. 26. Root, R. R., and K. M. Ragsdell, ‘‘Computational Enhancements to the Method of Multipliers,’’ ASME J. Mech. Des., 102(3), 517–523 (1980). 27. Gabriele, G. A., and K. M. Ragsdell, ‘‘Large Scale Nonlinear Programming Using the Generalized Reduced Gradient Method,’’ ASME J. Mech. Des., 102(3), 566– 573 (1980). 28. Katz, D. L., and E. R. Lady, Compressed Air Storage for Electric Power Generation, Ulrich’s Books, Ann Arbor, MI, 1976. 29. Ahrens, F. W., ‘‘The Design Optimization of Aquifer Reservoir-Based Compressed Air Energy Storage Systems,’’ Proc. 1978 Mech. Magnetic Energy Storage Contractors’ Rev. Meet., Luray, VA, CONF-781046, 1978, pp. 404–415.

APPENDIX

A

REVIEW OF LINEAR ALGEBRA

A.1

SET THEORY

A set is a well-defined collection of things. By well defined we mean that given any object it is possible to determine whether or not it belongs to the set. The set S ⫽ {x兩x ⭓ 0} defines the set of all nonnegative numbers. x ⫽ 2 is an element of the set S and is written as 2 僆 S (2 belongs to S). The union of two sets P and Q defines another set R such that R ⫽ P 傼 Q ⫽ {x兩x 僆 P or x 僆 Q or both}. The intersection of two sets, written P 艚 Q, defines a set R ⫽ {x兩x 僆 P and x 僆 Q}. P is a subset of Q, written P 傺 Q, if every element of P is in Q. Disjoint sets have no elements in common. If P and Q are disjoint sets, then x 僆 P implies x ⰻ Q, and vice versa. The empty set, denoted by ∅, is a set with no elements in it. A.2

VECTORS

A vector is an ordered set of real numbers. For instance, a ⫽ (a1, a2, . . . , an) is a vector of n elements or components. If a ⫽ (a1, a2, . . . , an) and b ⫽ (b1, b2, . . . , bn), then a ⫹ b ⫽ c ⫽ (a1 ⫹ b1, a2 ⫹ b2, . . . , an ⫹ bn) a ⫺ b ⫽ d ⫽ (a1 ⫺ b1, a2 ⫺ b2, . . . , an ⫺ bn) For any scalar ␣ positive or negative, Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

633

634


␣a ⫽ (␣a1, ␣a2, . . . , ␣an)

The vector O ⫽ (0, 0, . . . , 0) is called the null vector. The inner product of two vectors, written a 䡠 b or simply ab, is a number given by a1b1 ⫹ a2b2 ⫹ 䡠䡠䡠 ⫹ anbn For example, if a ⫽ (1, 2, 3) and b ⫽

冢冣

4 5 , then a 䡠 b ⫽ 4 ⫹ 10 ⫹ 18 ⫽ 6

32. A set of vectors a1, a2, . . . , an is linearly dependent if there exist scalars ␣1, ␣2, . . . , ␣n not all zero such that

冘␣a ⫽0 n

i

i

i⫽1

In this case, at least one vector can be written as a linear combination of the others. For example, a1 ⫽ ␭2a2 ⫹ ␭3a3 ⫹ 䡠䡠䡠 ⫹ ␭n an If a set of vectors is not dependent, then it must be independent. A vector space is the set of all n-component vectors. This is generally called the Euclidean n-space. A set of vectors is said to span a vector space V if every vector in V can be expressed as a linear combination of the vectors in that set. A basis for the vector space V is a set of linearly independent vectors that spans V.

A.3

MATRICES

A matrix A of size m ⫻ n is a rectangular array (table) of numbers with m rows and n columns. Example A.1

A ⫽ (2⫻3)

冋

1 4

2 5

3 6

册

is a matrix of two rows and three columns. The (i, j)th element of A, denoted by aij, is the element in the ith row and jth column of A. In this example, a12 ⫽ 2 while a23 ⫽ 6. In general an m ⫻ n matrix is written as

A.3

MATRICES

635

A ⫽ [aij] (m⫻n)

The elements aij for i ⫽ j are called the diagonal elements; while the aij for i ⫽ j are called the off-diagonal elements. The elements of each column of a matrix define a vector called a column vector. Similarly, each row of a matrix defines a row vector. In this example, the vectors a1 ⫽

冉冊 1 4

a2 ⫽

冉冊 2 5

a3 ⫽

冉冊 3 6

are the column vectors of the matrix A, while the vectors b1 ⫽ (1, 2, 3) and b2 ⫽ (4, 5, 6) are the row vectors of A. Thus a vector may be treated as a special matrix with just one row or one column. A matrix with an equal number of rows and columns is called a square matrix. The transpose of a matrix A ⫽ [aij], denoted by A⬘ or AT, is a matrix obtained by interchanging the rows and columns of A. In other words, AT ⫽ [aij⬘ ], where aij⬘ ⫽ aji. The transpose of A defined in this example is given by

冤冥

1 AT ⫽ 2 (3⫻2) 3

4 5 6

The matrix A is said to be symmetric if AT ⫽ A. The identity matrix, denoted by I, is a square matrix whose diagonal elements are all 1 and the off-diagonal elements are all zero. A matrix whose elements are all zero is called a null matrix. A.3.1

Matrix Operations

The sum or difference of two matrices A and B is a matrix C (written C ⫽ A Ⳳ B), where the elements of C are given by cij ⫽ aij Ⳳ bij For two matrices A and B the product AB is defined if and only if the number of columns of A is equal to the number of rows of B. If A is an m ⫻ n matrix and B is an n ⫻ r matrix, then the product AB ⫽ C is defined as a matrix whose size is m ⫻ r. The (i, j)th element of C is given by

冘a b n

cij ⫽

ik

k⫽1

kj

636


Example A.2

A ⫽ (2⫻3)

冋

册

1 2 3 4 5 6

AB ⫽ C ⫽ (2⫻2)

冋

冤冥

1 B ⫽ 3 (3⫻2) 5

22 28 49 64

册

2 4 6

Example A.3

Let A⫽

冋

册

冤冥

2 x⫽ 3 4

1 2 3 4 5 6

y ⫽ (2, 3)

Then Ax ⫽ b ⫽ (2⫻1)

冋册 20 47

while yA ⫽ d ⫽ (14, 19, 24) (1⫻3)

Note that b is a column vector and d is a row vector. For any scalar ␣, ␣A ⫽ [␣aij]

Matrix operations satisfy the following properties: 1. 2. 3. 4. 5. 6. 7.

(A ⫹ B) ⫹ C ⫽ A ⫹ (B ⫹ C) A⫹B⫽B⫹A (A ⫹ B)C ⫽ AC ⫹ BC (AB)C ⫽ A(BC) IA ⫽ AI ⫽ A (A ⫹ B)T ⫽ AT ⫹ B⬘ (AB)T ⫽ B⬘AT

In general, AB ⫽ BA.

A.3

A.3.2

MATRICES

637

Determinant of a Square Matrix

The determinant of a square matrix A, denoted by 兩A兩, is a number obtained by certain operations on the elements of A. If A is a 2 ⫻ 2 matrix, then 兩A兩 ⫽

冏

冏

a11 a12 ⫽ a11a22 ⫺ a12a21 a21 a22

If A is an n ⫻ n matrix, then

冘 a (⫺1) n

兩A兩 ⫽

兩Mi1兩

i⫹1

i1

i⫽1

where Mi1 is a submatrix obtained by deleting row i and column 1 of A. For example, if

冤

1 2 3 A ⫽ 4 5 6 (3⫻3) 7 8 9

冥

then 兩A兩 ⫽ 1

冏冏冏冏冏冏 5 6 2 3 2 3 ⫺4 ⫹7 8 9 8 9 5 6

⫽ (45 ⫺ 48) ⫺ 4(18 ⫺ 24) ⫹ 7(12 ⫺ 15) ⫽ 0

A matrix is said to be singular if its determinant is equal to zero. If 兩A兩 ⫽ 0, then A is called nonsingular. A.3.3

Inverse of a Matrix

For a nonsingular square matrix A, the inverse of A, denoted by A⫺1, is a nonsingular square matrix such that AA⫺1 ⫽ A⫺1A ⫽ I (identity matrix) The inverse matrix A⫺1 may be obtained by performing row operations on the original matrix A. The row operations consist of the following: 1. Multiply or divide any row by a number. 2. Multiply any row by a number and add it to another row.

638


To find the inverse of a matrix A, one starts by adjoining an identity matrix of similar size as [A]. By a sequence of row operations, A is reduced to I. This will reduce the original I matrix to A⫺1, since A⫺1[AI] ⫽ [IA⫺1] Example A.4

冋册 1 1

A⫽

1

⫺1

Since 兩A兩 ⫽ ⫺2, A is nonsingular, and hence A⫺1 exists. To compute A⫺1, start with the matrix (AI) ⫽

冋

1 1

册

1 兩兩 1 0 ⫺1 兩 0 1

Subtract row 1 from row 2:

冋

1 0

1 兩兩 1 0 ⫺2 兩 ⫺1 1

冋

1 兩兩 1 0 1 兩 –12 ⫺–12

册

Divide row 2 by ⫺2: 1 0

册

Subtract row 2 from row 1:

冋

1 0

册

–12 0 兩兩 –12 1 – 1 兩 2 ⫺–12

Thus, A⫺1 ⫽ Verify that AA⫺1 ⫽ A⫺1A ⫽ I.

冋册 –21 –12

–21 ⫺–12

A.3

A.3.4

MATRICES

639

Condition of a Matrix

The condition of the matrix is a measure of the numerical difficulties likely to arise when the matrix is used in calculations. Consider, for example, the quadratic function Q(x) ⫽ xTA⫺1x The matrix A is said to be ill-conditioned if small changes in x produce large changes in the value of Q. Ill-conditioning is often associated with A being very nearly singular. The condition number of a matrix as used in this book is given by K(A) ⫽

冏冏 ␭h ␭l

where ␭h and ␭l are the eigenvalues of greatest and smallest modulus, respectively. A matrix with large condition number is ill-conditioned; a matrix with condition number close to 1 is well-conditioned.

A.3.5

Sparse Matrix

A sparse matrix is one that has relatively few nonzero elements. For example, a 100 ⫻ 100 identity matrix will have 104 elements but only 100 of these (the diagonal elements) are nonzero; hence, this matrix has only 1 percent nonzero elements. To avoid storage of and arithmetic operations with the zero elements, sparse matrices are usually stored in packed form, operations involving matrix elements are carried out involving only the nonzero elements, and the operations are sequenced to generate as few new nonzero elements as practical. This extra logic constitutes an overhead that must always be balanced against the savings obtained by sparse-matrix methods. One simple packed form of storage is to store the nonzero elements in a linked list supplemented by a column pointer vector. Thus array elements are stored in a long vector consisting of a string of triplets (i, a, p). The entries of each triplet are associated with a specific nonzero element; i is the row index of the element, a is the value of the element, and j is the pointer that gives the location of the triplet associated with the next nonzero element in that row of the matrix. The column pointer vector is simply an n-component vector that gives the location of the triplet associated with first nonzero row element of each column of the matrix. For further discussion of sparse matrices and methods, the reader is directed to R. P. Tewarson, Sparse Matrices, Academic, New York, 1973.

640

A.4


QUADRATIC FORMS

A function of n variables ƒ(x1, x2, . . . , xn) is called a quadratic form if

冘冘 q x x ⫽ x Qx n

ƒ(x1, x2, . . . , xn) ⫽

n

T

ij i j

i⫽1 j⫽1

where Q(n⫻n) ⫽ [qij] and xT ⫽ (x1, x2, . . . , xn). Without any loss of generality, Q can always be assumed to be symmetric. Otherwise, Q may be replaced by the symmetric matrix (Q ⫹ Q⬘)/2 without changing the value of the quadratic form. Definitions 1. A matrix Q is positive definite if and only if the quadratic form xTQx ⬎ 0 for all x ⫽ 0. For example, Q⫽

冋

册

2 ⫺1 2

⫺1

is positive definite. 2. A matrix Q is positive semidefinite if and only if the quadratic form xTQx ⭓ 0 for all x and there exists an x ⫽ 0 such that xTQx ⫽ 0. For example, Q⫽

冋

1

⫺1

册

⫺1

1

is positive semidefinite. 3. A matrix Q is negative definite if and only if ⫺Q is positive definite. In other words, Q is negative definite when xTQx ⬍ 0 for all x ⫽ 0. For example, Q⫽

冋

⫺2

1

册

1

⫺3

is negative definite. 4. A matrix Q is negative semidefinite if ⫺Q is positive semidefinite. For example,

A.4

Q⫽

冋

⫺1

1

QUADRATIC FORMS

641

册

1

⫺1

is negative semidefinite. 5. A matrix Q is indefinite if xTQx is positive for some x and negative for some other x. For example, Q⫽

冋册 1 ⫺1 1 ⫺2

is indefinite. A.4.1

Principal Minor

If Q is an n ⫻ n matrix, then the principal minor of order k is a submatrix of size k ⫻ k obtained by deleting any n ⫺ k rows and their corresponding columns from the matrix Q. Example A.5

冤

1 Q⫽ 4 7

2 5 8

3 6 9

冥

Principal minors of order 1 are essentially the diagonal elements 1, 5, and 9. The principal minor of order 2 are the following (2 ⫻ 2) matrices:

冋册冋册 1 2 4 5

1 3 7 9

and

冋册 5 6 8 9

The principal minor of order 3 is the matrix Q itself. The determinant of a principal minor is called the principal determinant. For an n ⫻ n square matrix, there are 2n ⫺ 1 principal determinants in all. The leading principal minor of order k of an n ⫻ n matrix is obtained by deleting the last n ⫺ k rows and there corresponding columns. In Example A.5, the leading principal minor of order 1 is 1 (delete the last two rows and columns). The leading principal minor of order 2 is

冋册, while that of order 1 4

2 5

3 is the matrix Q itself. The number of leading principal determinants of an n ⫻ n matrix is n. There are some easier tests to determine whether a given matrix is positive definite, negative definite, positive semidefinite, negative semidefinite, or in-

642


definite. All these tests are valid only when the matrix is symmetric. [If the matrix Q is not symmetric, change Q to (Q ⫹ Q⬘)/2 and then apply the tests.] Tests for Positive-Definite Matrices 1. All diagonal elements must be positive. 2. All the leading principal determinants must be positive. Tests for Positive-Semidefinite Matrices 1. All diagonal elements are nonnegative. 2. All the principal determinants are nonnegative. Remarks 1. Note the major difference in the tests for positive-definite and positivesemidefinite matrices. For the positive-definite property, it is sufficient that the leading principal determinants are positive, while for positivesemidefinite property, we must test whether all the principal determinants are nonnegative. 2. To prove that a matrix is negative definite (negative semidefinite), test the negative of that matrix for positive definiteness (positive semidefiniteness). 3. A sufficient test for a matrix to be indefinite is that at least two of its diagonal elements are of the opposite sign.

A.4.2

Completing the Square

Another approach to determine whether a symmetric matrix Q is positive definite, negative definite, or indefinite is to rewrite the quadratic form xTQx as a sum of perfect squares. The property of the quadratic form can then be determined from the coefficients of the squared terms. Example A.6

Given the quadratic function q(x) ⫽ ax 21 ⫹ 2bx1x2 ⫹ cx 22 collect the terms in x1;

A.4

q(x) ⫽

QUADRATIC FORMS

643

1 2 2 (a x 1 ⫹ 2abx1x2) ⫹ cx 22 a

Next, add and subtract an appropriate term in x 22 so that the term in parentheses becomes a perfect square: q(x) ⫽

1 2 2 (a x 1 ⫹ 2abx1x2 ⫹ b2x 22 ⫺ b2x 22) ⫹ cx 22 a

q(x) ⫽

1 b2 (ax1 ⫹ bx2)2 ⫺ x 22 ⫹ cx 22 a a

Finally, grouping the coefficients of the x 22 term, we obtain q(x) ⫽

1 1 (ax1 ⫹ bx2)2 ⫹ (ac ⫺ b2)x 22 a a

Now, if 1/a ⬎ 0 and ac ⫺ b2 ⬎ 0, then clearly q(x) will be ⬎ 0 for any choice of x 僆 RN. Similarly, if 1/a ⬍ 0 and ac ⫺ b2 ⬎ 0, then q(x) will be negative for all choices of x. For any other combination of coefficient signs, q(x) will go through a sign change and hence will be indefinite. Completing the square can be viewed as a change of variables that results in a perfect square. For instance, if in the above example we let z1 ⫽ ax1 ⫹ bx2

z2 ⫽ x2

or x1 ⫽

1 1 (z1 ⫺ bx2) ⫽ (z1 ⫺ bz2) a a

x2 ⫽ z2

then we obtain q(x) ⫽ ax 21 ⫹ 2bx1x2 ⫹ cx 22

冉

⫽a ⫽

q(x) ⫽

冊冉

z1 ⫺ bz2 a

2

⫹ 2b

冊

z1 ⫺ bz2 z2 ⫹ cz 22 a

1 2 1 (z 1 ⫺ 2bz1z2 ⫹ b2z 22) ⫹ (2bz1z2 ⫺ 2b2z 22) ⫹ cz 22 a a 1 2 1 z ⫹ (ac ⫺ b2)z 22 a 1 a

It will be helpful to examine these operations in matrix form. Begin, once again, with the quadratic

644


冋册再冎 a b

q(x) ⫽ [x1, x2]

b c

x1 x2

and introduce the change of variables:

再冎冤冥再冎 1 ⫺b a a 0 1

x1 ⫽ x2

z1 z2

Accordingly, we obtain

冥冋册冤冥再冎

冤冤冥冤冥再冎冤冥再冎

1 0 a q(z) ⫽ [z1, z2] ⫺b 1 a 1 a ⫽ [z1, z2] ⫺b a

q(z) ⫽ [z1, z2]

a b b c

0

1 ⫺b a a 0 1

1 0 b b2 c⫺ a a

1

1 a

0

0

1 (ac ⫺ b2) a

z1 z2

z1 z2

z1 z2

Clearly, the process of completing the square is equivalent to introducing a change of variable that will reduce a general symmetric matrix to diagonal form. Alternatively, we can view these operations as equivalent to finding the factors D and P of a symmetric matrix Q, where D is a diagonal such that Q ⫽ PTDP Fortunately, this can easily be done using Gaussian elimination procedure (see Appendix C) or reduction to row echelon form, with some slight modifications. We illustrate the procedure by continuing with our example. Start with Q1 ⫽

冋册 a b b c

and store a and d1. Following the standard reduction procedure, divide the first row by a, the pivot element, and subtract b times the result from the second row:

A.4

Q2 ⫽

QUADRATIC FORMS

645

冤冥 1

b a

0

c⫺

b2 a

Next, the second row is divided by its first nonzero element, the pivot, and that element is stored as d2. In our simple example, we have arrived at the desired row eschelon form, and the process terminates. Furthermore, we observe that with the final pivot removed,

冤冥 b a 0 1 1

Q3 ⫽

we have generated the permutation matrix P, that is, P ⫽ Q3 ⫽

冤冥 b a 1

1 0

Furthermore, the required diagonal matrix simply contains the stored pivot elements:

D⫽

冤

冥

a

0

0

1 (ac ⫺ b2) a

One can readily verify the decomposition by performing the matrix multiplication PTDP to produce Q. With this construction, determining whether a symmetric matrix is positive or negative definite consists of merely performing Gaussian elimination on the matrix and saving the pivot elements. If all of the pivot elements are positive, the matrix is positive definite; if all are negative, the matrix is negative definite; otherwise it is indefinite. Example A.7

Given Q⫽

冋

12

⫺2

⫺2

10

册

We investigate the definiteness of Q via the reduction process. Here d1 ⫽ 12, and the first reduced array is

646


冋

册

⫺–16

1 ⫺2

10

The final reduction gives

冋册 1 ⫺–16 29 0 –– 3

or

冋册 1 ⫺–16 0 1

29 with d2 ⫽ –– 3 . Accordingly, Q has been factored to

Q⫽

冋册冋 1

⫺– 1 6

0 1

12 0 29 0 –– 3

册冋册 1 ⫺–16 0 1

Since the two diagonal elements are positive, Q is positive definite. A.5

CONVEX SETS

A set S is said to be a convex set if for any two points in the set the line joining those two points is also in the set. Mathematically, S is a convex set if for any two vectors x (1) and x (2) in S the vector x ⫽ ␭x (1) ⫹ (1 ⫺ ␭)x (2) is also in S for any number ␭ between 0 and 1. Examples

Figures A.1 and A.2 represent convex sets, while Figure A.3 is not a convex set. Theorem A.1 The set of all feasible solutions to a linear programming problem is a convex set. Theorem A.2 The intersection of convex sets is a convex set (Figure A.4). Theorem A.3 The union of convex sets is not necessarily a convex set (Figure A.4).

Figure A.1. Convex set.

Figure A.2. Convex set.

Figure A.3. Nonconvex set.

A.5

CONVEX SETS

647

Figure A.4. Intersection and union of convex sets.

Definitions A convex combination of vectors x (1), x (2), . . . , x (k) is a vector x such that x ⫽ ␭1x (1) ⫹␭2x (2) ⫹ 䡠䡠䡠 ⫹ ␭k x (k) ␭1 ⫹ ␭2 ⫹ 䡠䡠䡠 ⫹ ␭k ⫽ 1 ␭i ⭓ 0

for i ⫽ 1, 2, . . . , k

An extreme point or vertex of a convex set is a point in the set that cannot be expressed as the midpoint of any two points in the set. For example, consider the convex set S ⫽ {(x1, x2)兩0 ⭐ x1 ⭐ 2, 0 ⭐ x2 ⭐ 2}. This set has four extreme points given by (0, 0), (0, 2), (2, 0), and (2, 2). A hyperplane is the set of all points x satisfying cx ⫽ z for a given vector c ⫽ 0 and scalar z. The vector c is called the normal to the hyperplane. For example, H ⫽ {(x1, x2, x3)兩2x1 ⫺ 3x2 ⫹ x3 ⫽ 5} is a hyperplane. A half space is the set of all points x satisfying cx ⭐ z or cx ⭓ z for a given vector c ⫽ 0 and scalar z. Theorem A.4 A hyperplane is a convex set. Theorem A.5 A half space is a convex set.

APPENDIX

B

CONVEX AND CONCAVE FUNCTIONS Convex Function

A function of n variables ƒ(x) defined on a convex set D is said to be a convex function if and only if for any two points x (1) and x (2) 僆 D and 0 ⭐ ␭ ⭐ 1, ƒ[␭ x (1) ⫹ (1 ⫺ ␭)x (2)] ⭐ ␭ƒ(x (1)) ⫹ (1 ⫺ ␭)ƒ(x (2)) Figure B.1 illustrates the definition of a convex function of a single variable. Properties of Convex Functions

1. The chord joining any two points on the curve always falls entirely on or above the curve between those two points. 2. The slope or first derivative of ƒ(x) is increasing or at least nondecreasing as x increases. 3. The second derivative of ƒ(x) is always nonnegative for all x in the interval. 4. The linear approximation of ƒ(x) at any point in the interval always underestimates the true function value. 5. For a convex function, a local minimum is always a global minimum. Figure B.2 illustrates property 4. The linear approximation of ƒ at the point x 0, denoted by ƒ˜ (x; x 0), is obtained by ignoring the second and other higher order terms in the Taylor series expansion ˜ ƒ(x; x 0) ⫽ ƒ(x 0) ⫹ ⵜƒ(x 0)(x ⫺ x 0)

For a convex function, property 4 implies that

648


CONVEX AND CONCAVE FUNCTIONS

649

Figure B.1. Convex function.

ƒ(x) ⭓ ƒ(x 0) ⫹ ⵜƒ(x 0)(x ⫺ x 0)

for all x

The gradient of a function ƒ(x1, . . . , xn) is given by ⵜƒ(x1, . . . , xn) ⫽

冋

␦ƒ ␦ƒ ␦ƒ , ,..., ␦x1 ␦x2 ␦xn

册

The Hessian matrix of a function ƒ(x1, . . . , xn) is an n ⫻ n symmetric matrix given by Hƒ(x1, . . . , xn) ⫽

冋册

␦2ƒ ⫽ ⵜ2ƒ ␦xi ␦xj

Figure B.2. Linear approximation of a convex function.

650

CONVEX AND CONCAVE FUNCTIONS

Test for Convexity of a Function. A function ƒ is convex function if the

Hessian matrix of ƒ is positive definite or positive semidefinite for all values of x1, . . . , xn. Concave Function

A function ƒ(x) is a concave function in D if and only if ⫺ƒ(x) is a convex function in D. Test for Concavity of a Function. A function ƒ is concave if the Hessian

matrix of ƒ is negative definite or negative semidefinite for all values of x1, . . . , xn. Example B.1

ƒ(x1, x2, x3) ⫽ 3x 12 ⫹ 2x 22 ⫹ x 32 ⫺ 2x1x2 ⫺ 2x1x3 ⫹ 2x2 x3 ⫺ 6x1 ⫺ 4x2 ⫺ 2x3 ⵜf(x1, x2, x3) ⫽

冢冤

6x1 ⫺ 2x2 ⫺ 2x3 ⫺ 6 4x2 ⫺ 2x1 ⫹ 2x3 ⫺ 4 2x3 ⫺ 2x1 ⫹ 2x2 ⫺ 2

6 ⫺2 4 Hƒ(x1, x2, x3) ⫽ ⫺2 ⫺2 2

冣

冥

⫺2

2 2

To show that ƒ is a convex function, we test for whether H is positive definite or positive semidefinite. Note that: 1. H is symmetric. 2. All diagonal elements are positive. 3. The leading principal determinants are 兩6兩 ⬎ 0

冏

冏

6 ⫺2 ⫽ 20 ⬎ 0 4

⫺2

兩Hƒ兩 ⫽ 16 ⬎ 0

Hence H is a positive-definite matrix, which implies ƒ is a convex function. (As a matter of fact, when Hƒ is positive definite, ƒ is said to be strictly convex with a unique minimum point.)

APPENDIX

C

GAUSS–JORDAN ELIMINATION SCHEME Systems of linear equations can be solved by the classical Gauss–Jordan elimination procedure. In this Appendix, a brief review of this procedure is given through an example. Example C.1

Consider the system of two equations in five unknowns denoted by x1 ⫺ 2x2 ⫹ x3 ⫺ 4x4 ⫹ 2x5 ⫽ 2

(C.1)

x1 ⫺ x2 ⫺ x3 ⫺ 3x4 ⫺ x5 ⫽ 4

(C.2)

(S1)

Since there are more unknowns than equations, this system will have more than one solution. The collection of all possible solutions to the system is called the solution set. Definition Two systems of equations are said to be equivalent if both systems have the same solution set. In other words, a solution to one system is automatically a solution to the other system, and vice versa. The method of solving a system of equations is to get an equivalent system that is easy to solve. By solving the simple system, we simultaneously get the solutions to the original system. Engineering Optimization: Methods and Applications, Second Edition. A. Ravindran, K. M. Ragsdell and G. V. Reklaitis © 2006 John Wiley & Sons, Inc. ISBN: 978-0-471-55814-9

651

652

GAUSS–JORDAN ELIMINATION SCHEME

There are two types of elementary row operations that can be used to obtain equivalent systems: 1. Multiply any equation in the system by a positive or negative number. 2. Add to any equation a constant multiple (positive, negative, or zero) of any other equation in the system. An equivalent system to system (S1) can be obtained by multiplying Eq. (C.1) by ⫺1 and adding the result to Eq. (C.2) as follows: x1 ⫺ 2x2 ⫹ x3 ⫺ 4x4 ⫹ 2x5 ⫽ 2

(C.3)

x2 ⫺ 2x3 ⫹ x4 ⫺ 3x5 ⫽ 2

(C.4)

(S2)

From system (S2), another equivalent system can be obtained by multiplying Eq. (C.4) by 2 and adding it to Eq. (C.3). This gives system (S3): x1 ⫺ 3x3 ⫺ 2x4 ⫺ 4x5 ⫽ 6

(C.5)

x2 ⫺ 2x3 ⫹ x4 ⫺ 3x5 ⫽ 2

(C.6)

(S3)

Since systems (S1)–(S3) are equivalent, a solution to one automatically gives a solution to the other two. In this case, it is easy to write out all the possible solutions to system (S3). For example, setting x3 ⫽ x4 ⫽ x5 ⫽ 0 gives x1 ⫽ 6, x2 ⫽ 2, which is a solution to all three systems. Other solutions to system (S3) may be obtained by choosing arbitrary values for x3, x4, x5 and finding the corresponding values of x1 and x2 from Eqs. (C.5) and (C.6). All these solutions are solutions to the original system. Systems like (S3) are called canonical systems or systems in row echelon form.

AUTHOR INDEX

Abadie, J., 296, 373, 419, 427, 442, 476 Afimiwala, K. A., 280 Ahluwalia, R. K., 624, 627, 632 Ahrens, F. W., 621, 626, 628, 632 Albright, C., 493, 540 Arbel, A., 203, 207 Archetti, F., 587 Armarego, E. J. A., 23, 31, 511 Arthur, J. L., 514, 520–521 Avriel, M., 139, 282, 295 Baker, T. E., 355 Balas, E., 609 Bandler, J. W., 95 Bard, Y., 117, 463 Barnes, J. W., 168 Bartholomew-Biggs, M. C., 470 Bartle, R., 71 Bazaraa, M. S., 230, 253–254, 373, 411, 442 Beale, E. M. L., 122, 373, 463 Beck, P. A., 558 Beightler, C., 419 Beilby, M. H., 520 Benjamin, D. R., 553 Berlau, D. H., 595 Berna, T. J., 470–471, 474, 559 Bertsekas, D. P., 281, 297 Best, M. J., 293 Bett, K. E., 30, 595 Biegler, L. T., 443, 474, 595 Biggs, M. C., 470 Biles, W. E., 314, 554, 596 Bixby, R. E., 178, 190, 204–205, 207 Boland, W. R., 122 Borowski, N., 331 Bowler, A. T., 293 Box, G. E. P., 86, 92 Box, M. J., 69, 86, 313–314, 321, 331, 533, 563 Bracken, J., 280 Breiner, M., 540 Brent, R. P., 69, 97, 138

Brian, R., 540 Brooke, A., 179, 206, 540 Brooks, R., 248–249, 254 Brooks, S. H., 322, 533 Brown, R. H., 23, 31, 511 Broyden, G. G., 126 Buntin, R. R., 632 Burbridge, J. J., 520 Cabot, A. V., 588 Cantrell, J. W., 121 Carpenter, W. C., 131 Carpentier, J., 419, 426–427 Carroll, C. W., 280 Carter, W. J., 26–27 Casares Long, J. J., 331 Cauchy, A., 109 Chamberlain, R. M., 476 Chao, N. H., 471 Charnes, A., 519–520 Chiu, H. H., 632 Chung, S. F., 597 Clasen, R. J., 373 Clayton, E. R., 520 Coggins, G. F., 123, 138 Collatz, L., 541 Colville, A. R., 141, 533 Cook, W. D., 520 Cooper, L., 86 Cooper, W. W., 519–520 Cottle, R. W., 166, 506 Courant, R., 280 Crisp, R. M., Jr., 168 Crowder, H. P., 122, 140, 532 Cullinan-James, C., 183, 207 Czyzyk, J., 206 Dane, C. W., 520 Dantzig, G. B., 27, 31, 163, 189, 207, 373, 506 Datseris, P., 328, 331 Davidon, W. C., 124, 127


653

654

AUTHOR INDEX

Davies, D., 71, 137, 310, 321, 331 Davison, E. J., 122 Davison, W. R., 632 Deckro, R. J., 520 Dembo, R. S., 140, 532, 534 DeSilva, B. M. E., 280 Devroye, L. P., 332 Dixon, L. C. W., 139–140, 331, 476–477, 596 Doering, F. J., 326–327, 331 Dolan, E. D., 540 Draper, N. R., 92 Driebeek, N. J., 172 Drud, A., 442 Dupuis, T. M., 30 Ealey, L. A., 31 Eason, E. D., 134, 532–533 Eduljee, H. E., 615 El Agizy, 373 Embury, M. C., 594 Emery, F. E., 96 Engquist, M., 373 Everett, H., 237, 245, 254, 261, 283 Fang, S. C., 203–204, 207 Fattler, J. E., 596 Felt, A., 206 Feng, T. T., 419 Fenton, R. G., 134, 533 Fenvers, S. J., 476 Fiacco, A.V., 245, 278, 280, 282, 296 Fletcher, R., 85, 106, 118, 124, 126, 137– 138, 278, 281, 293, 296–297, 442, 476, 533 Fourer, R., 178–179, 206, 540 Fox, R. L., 280 Francis, R. L., 588 Frank, N., 339 Franz, J., 281, 296 Freedman, H. T., 520 Freudenstein, F., 301, 331 Fried, I., 118 Furtan, H., 498 Fylstra, D., 443 Gabriele, G. A., 123, 138, 281, 294, 296, 430–431, 632 Gaddy, J. L., 325–326, 331 Gaines, L. D., 331 Gallagher, R. H., 295, 540

Gannibandara, K., 596 Garcia-Palomares, U. M., 464–465 Gass, S. I., 174, 183 Gay, D. M., 206, 540 Geoffrion, A. M., 248–249, 254, 589, 596 Gere, J., 31 Gill, P. E., 106, 130, 286, 293, 295, 321, 331, 475 Glover, F., 484 Goldfarb, D., 126–127, 139 Gonzalez, S., 133 Gottfried, B., 595 Gottfried, X. X., 289 Goulcher, R., 331 Gould, F. J., 286 Grant, G. N. C., 280 Grantham, W. J., 30 Graves, G. W., 352 Graves, R. L., 373, 442 Gray, P., 183, 207 Greenberg, H. J., 373, 442 Greenstadt, J. L., 463 Gregory, J. W., 207 Griffith, F. E., 533, 538 Griffith, R. E., 352 Grossman, T. A., 493, 540 Grossmann, I. E., 443 Gupta, A., 581 Gupta, B. D., 280 Gupta, O. K., 493, 518 Gupta, S. K., 518 Guthrie, K. M., 596, 617 Hag, E. J., 419 Haimes, Y. Y., 520 Hall, W. A., 520 Hamilton, J. F., 140 Han, S. P., 127, 244, 464–465, 476 Hanline, D. L., 603, 631 Hans-Tijan, T. Y., 442 Harrald, J., 520 Haverly, C. A., 358, 373 Hayt, W. H., 31 Herbert, J. E., 520 Hesse, R., 600 Hestenes, M. R., 118, 283, 293 Heuckroth, M. W., 331 Hext, G. R., 87 Himmelblau, D. M., 69, 131, 280, 293, 312 Himsworth, F. R., 87 Holmes, R. L., 280, 541 Hooke, R., 93

AUTHOR INDEX

Howe, S., 297 Huard, P., 282 Hughes, R. R., 474, 554, 595 Hutchinson, D., 92 Hutchinson, H. P., 330 Hutchinson, J. J., 540 Hutchison, H. P, 553 Ignizio, J. P., 518 Ijiri, Y., 520 Jaakola, T. H. I., 329, 331 Jain, A., 442, 632 Jarvis, R. A., 331 Jeeves, T. A., 93 Jen, F. C., 11 Johnson, R. C., 419 Johnson, S. M., 31 Kamgnia, E. R., 139 Karmarkar, N. K., 180, 189–190, 194, 203, 206 Kartsounes, G. T., 621, 626 Karwan, K. R., 520 Katz, D. L., 627 Kawaratoni, T. K., 373 Keefer, D. L., 289, 595 Keeney, R. L., 518 Kemmerly, J. E., 31 Kendrick, D., 206, 540 Kernighan, B. W., 206, 540 Khachian, L. G., 180, 206 Kiefer, J., 49 Kilkas, A. C., 553 Kim, C. S., 621, 626 Kim, N. H., 373 King, J. C., 595 Klingman, W. R., 312 Knopf, F. C., 443, 596, 603–605 Kort, B. W., 286 Kowalik, J. S., 71, 139, 282 Krarup, J., 166 Krolak, P., 86 Kvalie, 335 Lady, E. R., 627 LaFrance, L. J., 127 Lagrange, 26–27 Lasdon, L. S., 237, 254, 280, 358, 373, 427, 442–443, 558, 632

655

Lavi, A., 540 Lawrence, K. D., 520 Lee, H., 499, 507 Lee, S. M., 517, 520 Lee, T. W., 331 Lemke, C. E., 499, 503 Leon, A., 533 Lerro, A. J., 520 Lessard, R. D., 632 Lieder, W. D., 554 Liepelt, M., 296 Lill, S., 280 Locke, M. H., 476, 559, 595 Lootsma, F. A., 137, 140, 278–280, 282, 294, 331, 476, 538, 596 Lu, J. Y., 373 Luenberger, D. G., 426, 595 Lustig, I. J., 204, 207 Luus, R., 329, 331, 601 Mah, R. S. H., 330–331 Mangasarian, O. L., 114, 140–141, 233, 236, 244, 253–254, 297, 464–465, 476–477 Maret, M. W., 520 Marquardt, D. W., 116, 128 Marsten, R. E., 178, 190, 203–204, 207 Martens, X. M., 572 Martin, D. L., 331 Mayne, D. Q., 476 Mayne, R. W., 531 McCann, A. P., 278 McCarl, B. A., 498 McCloskey, J. F., 31 McCormick, G. P., 127, 238, 254, 278, 280, 282, 296, 402, 541, 588 McDonald, P. A., 95 McLane, M., 554, 596 McMillan, C., Jr., 484, 493 Mead, R., 90–91, 312 Meador, N. C., 520 Meadows, J. C., 520 Meeraus, A., 206, 540 Meinardus, G., 541 Metzler, J., 118 Meyer, R. R., 140–141, 297, 476–477 Miele, A., 121, 133, 335 Miller, C. E., 359 Mockus, J., 331 Moe, J., 280 Moore, L. J., 520 More, J. J., 179, 206

656

AUTHOR INDEX

Moree`, J. J., 540 Moskowitz, H., 498, 521 Motard, R. L., 330 Mott, Jr., T. H., 520 Mulvey, J. M., 140, 532 Mungal, P., 632 Munson, T. S., 540 Murray, W., 85, 138, 140, 279, 293, 295, 297, 321, 331, 477 Murtagh, B. A., 428–429, 442 Murty, K. G., 207, 484, 493 Musa, A. A., 521 Muthusubramanyam, M., 520 Mylander, W. C., 280, 541 Nash, S. G., 530 Nelder, J. A., 90–91, 312 Nemhauser, G. L., 519 Newell, J. S., 293 Nicely, R., 520 Niehaus, R. J., 519 Nocedal, J., 137, 140, 281, 296 O’Hagan, M. O., 96 Okos, M. R., 443, 596, 603–604 Osborne, M. R., 71 Owen, J. H., 206 Palacios-Gomez, F. E., 352, 355, 358, 373 Parker, A. L., 595 Parkinson, J. M., 92 Pegden, C. D., 596 Pegels, C. C., 30 Philipson, R. H., 22, 31, 519 Phillips, D. T., 31, 71, 75, 177, 186, 206, 297, 484, 518 Phua, K. H., 127, 131, 140 Plane, D. R., 484, 493 Polak, E., 121 Pollack, H. W., 554 Popper, H., 632 Powell, M. J. D., 58, 93, 97, 117, 122, 124, 126–127, 140, 283, 464–466, 470–471, 477, 540 Pritsker, A. A. B., 595–596 Puthenpura, S., 203–204, 207 Ragsdell, K. M., 26–27, 123, 138, 140, 294, 296–297, 430–431, 443, 531, 541, 557, 595–596, 621, 632 Raiffa, H., 518 Rampla, G., 540 Rao, S. S., 280

Raphson, J., 61 Rardin, R. L., 484 Ratner, M. W., 442, 632 Ravindran, A., 22, 31, 71, 75, 177, 186, 206, 484, 493, 499, 507, 514, 518– 519, 520–521, 581, 603, 631 Rawlinson, J. S., 595 Reeves, C. M., 118 Reklaitis, G.V., 282, 331, 419, 443, 477, 581, 594–596, 603–605, 619, 631 Ribiere, G., 121 Rijckaert, H., 280 Rijckaert, M. J., 295, 509–511, 521–522, 572 Ritter, K., 293 Robinson, S. M., 140–141, 297, 476–477 Rockafellar, R. T., 283, 293 Root, R. R., 127, 138, 286, 289, 294, 296– 297, 557, 595–596, 632 Rosen, E. M., 595 Rosen, J. B., 537, 601 Rosenbrock, H. H., 96, 282 Rowlinson, J. S., 30 Rustagi, J. S., 442 Ryan, D. M., 280

Saaty, T. L., 538 Saha, J., 183 Salkin, H. M., 183 Saltzman, M., 207, 493 Salvendy, G., 166, 595 Sandgren, E., 280, 296, 534–536 Sargent, R. W. H., 131, 309, 581 Sarkar, S., 373 Sarma, P. V. L. N., 619, 631–632 Sauer, R. N., 598 Saunders, M. A., 140, 428–429, 442, 477 Saville, G., 30, 595 Saxena, U., 521 Schittkowski, K., 293, 296, 537, 541 Schniederjan, M. J., 517 Schrack, G., 327 Schuldt, S. B., 283 Schuler, A. T., 520 Schumer, M. A., 331 Scribner, B. F., 595 Sebastian, D. J., 131 Seider, W. D., 330 Sengupta, S., 521 Shah, J.V., 588, 596 Shanno, D. F., 122, 126–127, 131, 140, 178, 207 Sharda, R., 493, 539–540 Sharma, A., 626–627, 632

AUTHOR INDEX

Sherali, D., 254, 373 Sherwood, T. K., 141, 144, 589 Shetty, C. M., 254, 373, 411, 442 Shigley, J. E., 14 Shin, W., 521 Siddall, J. N., 263, 280 Sin, Y. T., 596 Sinha, K. C., 520 Sisser, F. S., 563–564 Smeers, Y., 442 Smith, B. D., 632 Smith, E. A., 131 Smith, H.V., 352 Smith, S., 442 Solberg, J. J., 71, 75, 177, 186, 206, 484, 518, 596 Sood, M. K., 595–596 Sorensen, D. C., 140 Spang, H. A., 322 Spedicato, E., 476–477, 596 Spendley, W., 87 Squires, R. G., 595, 605 Staha, R. L., 282 Steiglitz, K., 331 Stewart, G. W., 130 Stewart, P. B., 632 Stewart, R. A., 352, 533, 538 Stiefel, E., 118 Subramanian, 207 Suzuki, S., 537, 601 Swain, J. J., 554, 596 Swann, W. H., 46, 71, 96, 137, 321, 331 Szego, G. P., 331, 476–477, 587, 596 Taguchi, G., 27–28, 31 Taylor, B. W., 520 Tewarson, R. P., 639 Timoshenko, S., 31 Toint, Ph. L., 140 Topkis, D. M., 385, 387, 442 Trefethen, F. N., 31 Treybal, R. E., 632 Trivedi, V. M., 521

657

Ullman, R. J., 373

Vienott, A. F., Jr., 385, 387, 442 Vincent, T. L., 30 Vogl, W., 540 Vosburgh, K. G., 632

Wait, R., 330 Wallace, W. A., 520 Waren, A. D., 355, 373, 442, 632 Watson, G. A., 476 Watters, L., 483 Waven, A., 443 Webster, H. H., 520 Westerberg, A. W., 330, 443, 476, 559, 588, 595–596 Wetterling, W., 541 White, J. B., 520 White, W. B., 22 Wier, C. E., 628 Wilde, D. J., 295, 419, 477, 596 Wilson, R. B., 463 Winkofsky, E. P., 520 Winter, P., 309 Wolfe, P., 122, 339, 345, 373, 385, 402, 442, 499 Wong, P., 122 Woods, J. M., 567, 581 Woolsey, E., 484 Woolsey, L. A., 519 Wright, M. H., 138, 140, 477 Wright, S. J., 140, 179, 206, 281, 296

Zanakis, S. H., 518, 520 Zangwill, W. I., 97, 138, 280, 345, 405, 441–442 Zeleny, M., 30, 518 Zhang, J., 355, 373 Zienkiewicz, O. C., 295, 540 Zilinskas, A., 331 Zoutendijk, G., 378, 441

SUBJECT INDEX

Academic planning, 517 ACDPAC: method of multipliers program, 293 Active constraint set, 380, 383 jamming, 385 Adjacent basic feasible solution, 164 Algorithm: robustness, 532 testing, 532 Alkylation process optimization, 598 Almost complementary solution, 503 Alternate: optima, 171 optimal solution, 156, 161 AMPL, 179, 531, 540 Applications of optimization, 6, 603, 610, 621 Artificial variables, 174 Ascent direction, 82 Augmented Lagrangian method, 282

Bala’s implicit enumeration method, 609 Barrier methods, 261 Basic feasible solution, 162 Basic solution, 162 Basic variables, 162 Basis, 634 Batch / semicontinuous process design, 432 Bethlehem Steel Corporation equation generator, 553 BIAS: application example, 627 comparison with other methods, 536 method of multipliers program, 291, 558 Bisection method, 63, 69 Bolzano search, 63 Boundary points, 41 Bracket operator penalty function, 264, 269 Branch-and-bound algorithm, 484 Broyden-Fletcher-Goldfarb-Shanno method, 126, 427, 465

Canonical form, 162 Canonical reduction, 162 Canonical systems, 652 Case studies, 603 Cauchy’s method, 110, 113 Central difference formula, 130 Chemical equilibrium problem, 22 Classical mechanics, 26 Coal blending application, 603 Coefficient matrix, 159 Column vector, 635 Colville study, 533 Complementary pivot method, 499, 502 Complementary pivot problems, 499 Complementary problem, 500 Complementary slackness condition, 227 Complementary slackness theorem, 187 Complementary solution, 502 Completing the square, 642 Complex method, 312, 320 algorithm outline, 314 COMPUTE II: penalty function program, 280 Computer cycling, 174 Concave: function, 650 Condition: of matrix, 639 number, of matrix, 639 Conjugate: directions, 100, 102 method, 310 gradient methods, 109, 117, 146. See also Unconstrained multivariable optimization Constrained derivative conditions, 419 Constrained derivative formula, 409 Constrained direction generation methods, 378 convex simplex method, 389 feasible directions, 378 generalized reduced gradient method, 406


659

660

SUBJECT INDEX

Constrained direction generation methods (continued ) reduced gradient method, 399 see also Reduced gradient Constrained direct search: complex method, 312 random search methods, 322 Constrained optimality criteria, 218 Constrained quadratic approximation methods, 450 direct, 451 for Lagrangian function, 456 and Newton’s method, 461 use of penalty function, 462 variable metric methods, 464 Constrained regression problem, 497 Constrained search methods, 305 Constrained variable metric method: algorithm for:, 464 Broyden-Flecther-Goldfarb-Shanno updating formula, 465 comparison with generalized reduced gradient method, 471 constraint inconsistency, 470 convergence results, 470 modification of H(t), 471 problem scaling, 470 Constraint elimination, 307 Constraint qualification, 230 Constraints, 22 Continuous functions, 33 Convergence, 106 Convergence rate, 107 Convex: combination, 647 function, 83, 648 set, 646 simplex method, 389, 507 algorithm for, 396 basis selection rules, 399 complementary slackness condition, 395 convergence acceleration, 405 Kuhn-Tucker convergence, 399 nondegeneracy assumption, 399 nonnegativity condition, 395 relative cost factor, 390 target point, 396 Convexity, 567 CPLEX, 178, 493 Cubic search, 65, 69 Cycling, 174

DAVID: penalty function program, 281 Davidon-Fletcher-Powell formula, 465 Davidon-Fletcher-Powell method, 124, 374 Decision vector, 159 Decomposition: method, application example, 621 procedures, 309 Degeneracy, 174 Dependent variables, 162 Descent: direction, 82, 338, 379 property, 110 Determinant, of matrix, 637 Differential algorithm, 419 Direct Linear approximation problem, 346 Direct methods, for unconstrained multivariable optimization, 85 Direct quadratic approximation, 451 algorithm for, 452 Direct search methods: constrained, 305 equality constraints, 306 Discontinuous functions, 33 Dual, 184 Duality gap, 201 Duality theory:, 184 complementary slackness theorem, 187 duality gap, 201 main duality theorem, 187 optimality criterion theorem, 187 symmetric dual, 185 weak duality theorem, 186, 201 Dual simplex method, 188 Dyanmic question, in optimization, 35, 79 Eason and Fenton study, 533 Economic machining problem, 511 Economic order quantity, 43–44 Electrical resistance network, 22 Engineering case studies, 603 Equality constraints, 28 Equation generators, 553 Equation ordering, 309 Euclidiean n-space, 634 Evolutionary optimization, 86 Excel Solver, 177, 182, 440, 493, 530 Execution strategies, 580 for equation-oriented models, 580 for global optimization, 587 for simulation models, 581 implicit constraints, 582

SUBJECT INDEX

internal numerical procedures, 582 multiple-response surface, 587 two-level, 581–582 sequential approaches, 580 Exhaustive search, 54 Extended parallel subspace property, 104 Exterior penalty function, 261. See also Penalty, function Extreme point, 647

Facility location problem, 603 Fathoming, in mixed integer programs, 489, 491 Feasible direction, 379 Feasible direction methods, 378 conditions for use of, 387 convergence of, 387 e-perturbation method, 385 linear programming subproblem, 380 Topkis-Veinott variant, 385 Feasible region, 154, 161 Feasible solution, 154, 161 Feasible starting point generation: infeasibility minimization, 576 procedures for, 576 sequential constraint minimization, 576 see also Starting point generation Fixed charge problem, 482 Fixed-parameter penalty methods, 282 Flash separator model, 549 Fletcher-Reeves formula, 427 Fletcher-Reeves method, 120 Fluid-processing facility problem, 583 Forest management, 517 Forward difference approximation, 130 Frank-Wolfe algorithm, 339 convergence of, 345 Front line systems, 531 Fuel allocation, in power plants, 323

GAMS, 179, 531, 540 Gantt chart, 435 GAPFPR: method of multipliers program, 293 Gaussian elimination procedure, 644 Gauss-Jordan elimination scheme, 162, 651 Generalization of convex functions, 249 Generalized geometric programming, 581 Generalized Lagrange multiplier method, 237, 245

661

Generalized reduced gradient: active constraint strategy, 421 algorithm, 410, 415 application example, 618, 627 approximate Newton calculations, 427 computational enhancements, 427 constraint iterations, 414 convergence acceleration techniques, 426 convergence conditions, 426 convergence rate, 426 inside-out search strategy, 418 linear element treatment, 428 matrix inverse avoidance, 428 Newton’s method, 414 nondegenerate nonsingular basis, 426 nonlinear variables, 429 numerical derivatives, 429 superbasic variables, 429 treatment: of bounds, 419 of inequalities, 420 Generalized reduced gradient method, 406 comparison with other methods, 539 implicit variable elimination, 406 Geometric programming problems, 572 Global maximum, 36, 80 Global minimum, 36, 80, 83 Global optimization strategies, 587 for structured problems, 588 multistart, 587 Goal programming:, 205, 481, 507, 581 applications, 517 goal constraint, 508 partitioning algorithm, 514, 581 preemptive priority, 509 Golden section search, 51, 54–55, 69 Gradient, 649 method, 128. See also Generalized reduced gradient; Reduced gradient Gradient-based methods, 108. See also Generalized reduced gradient; Reduced gradient GRG, comparison with other methods, 539 GRG2, 440 Gulf Oil Corporation, 280

Half-space, 647 Hessian matrix, 80, 649–650 Himmelblau’s function, 79 Hooke-Jeeves pattern search, 93 Huard’s method of centers, 282

662

SUBJECT INDEX

Hydroelectric power systems, planning of, 151 Hyperplane, 647 Identity matrix, 635 Implicit variable elimination, 410 Indefinite matrix, 641–642, 645 Independent variables, 162 Inequality constraints, 28 Inequality penalty function, 269 Infeasibility, 161 Infinite barrier penalty function, 262 Inflection point, 38 Inner product rule, 166 Integer linear programs, 581 Integer programming, 205, 481 0–1 problem solutions, 493 0–1 variables, 604 binary, 483 branch-and-bound algorithm, 484 formulation, 482, 604 nonlinear 0–1 integer problems, 483 problem formulation guidelines, 492 problem solution, 484 Interior penalty function, 261 Interior point methods, 180, 189 affine scaling algorithm, 191 duality gap, 201 Karmarkar’s algorithm, 190 projection matrix, 196 projective scaling algorithm, 190 scaling matrix, 196 International Business Machines Corporation, integer program formulation, 492 Interval halving, 48–50, 54 Inventory control, 42–44 Inverse penalty function, 264, 273 Isolated local minimum, 238 Jamming, in direction-generating subproblems, 385 Karmarkar’s algorithm See Interior point methods Kuhn-Tucker conditions, 225, 498–499 complementary slackness, 227 constraint qualification, 230 for optimality, 501 interpretation of, 228 necessity theorem, 230 sufficiency theorem, 233

Kuhn-Tucker multiplier, 239 Kuhn-Tucker point, 233–234, 239 and penalty functions, 282 Kuhn-Tucker problem, 226 Kuhn-Tucker theorems, 230

Lagrange function, and penalty functions, 283 Lagrange multiplier, 219 and quadratic approximation, 457 Lagrangian function, 220, 236 and quadratic approximation, 456 Lagrangian necessary conditions, and quadratic approximation, 457 Land use planning, 517 Leading principal minor, 641 LINDO, 181, 493 Line search See Univariate search Linear approximation, 407 Linear combination, 634 Linear convergence rate, 107 Linear dependence, 634 Linearization, of a function, 336 Linearization methods: for constrained problems, 336 cutting plane, 327 Frank-Wolfe algorithm, 339 separable programming, 359 successive linear programming, 337 Linearization point, 337 Linearly constrained nonlinear programming problems, 388–389 Linearly constrained problem, nonlinear objective function, 337 Linear programming, 149 applications of, 183 computer codes, 177 duality theory, 184 dual simplex method, 188 formulation, 149 graphical solution for, 154 interior point method, 180, 189 modeling languages, 178 sensitivity analysis, 180 simplex method, 163 software, 177 standard form, 158 two-phase simplex method, 176 Linear simplex method matrix form of, 389 Local maximum, 36, 80 Local minimum, 36, 80, 238 Log penalty function, 263, 270

SUBJECT INDEX

Main duality theorem, 187 Manpower planning, 517 MAP linearization program, 352 Marquardt’s compromise, 116 MATLAB, 531, 540 Matrices, 634 determinant, 637 difference, 635 product, 635 sum, 635 Matrix, 634 condition number, 639 condition of, 639 generator, 553 identity, 635 inverse, 637 leading principal minor, 641 operations, 635 principal determinant, 641 principal minor, 641 singularity, 637 sparse, 639 submatrix, 637, 641 symmetric, 635 transpose, 635 Maximum, of a function, 80 MEMGRD: penalty function program, 281 Memory gradient method, 121 Metal cutting, 22, 517 Method of centers, 282 Method of multipliers: characteristics of, 286 comparison with other methods, 539 computer programs for, 293 multiplier update rule, 283 penalty function, 283 scale of problem, 289 termination, 285 Metric, 123 Minimax strategy, 51 Minimum, of a function, 80 Minimum ratio rule, 167 MINOS:, 531 reduced gradient program, 429 Mixed integer program (MIP), 481 facility location problem, 603 solution of, 484 Model, 5 Model formulation: case study of, 603, 610, 621 equation-oriented models, 548 flash separator model, 549 levels of modeling, 544

663

procedure for, 543 response surface models, 548 simulation models, 548 types of models, 548 Model preparation: boundedness, 567 case study, 603, 610, 621 equation sequencing, 566 feasibility, 575 function transformation, 560 inconsistencies, 559 invalid arguments, 559 nondifferentiability, 559 numerical difficulties in, 555 partial derivative test, 568 problem feature analysis, 566 recursive scaling method, 558 redundancy elimination, 565 solution uniqueness, 570 variable test, 568 variable transformation, 562 Modified Newton’s method, 115 Modified simplex procedure, 90 Monotonic function, 34 Monotonicity, 34 MPS format, 553 Multimedia filter optimization, 599 Multimodal functions, 69 Multimodal objective function, 79 Multiple local minima, 570 Multiple minima, 572 Multistage compressor optimization, 340

Necessary conditions, for multi variable function, 82 Negative definite matrix, 640, 642, 645, 650 Negative semidefinite matrix, 640, 642, 650 NEOS, 179, 531 Newton-Raphson method, 61 Newton’s method: modified, 115 for unconstrained multivariable optimization, 111 Nonbasic variables, 162 Nonlinear curve fitting, 21 Nonlinear program codes, Schittkowski comparison study, 537 Nonlinear programming, application example, 610, 621. See also Unconstrained, multivariable optimization Nonlinear programming problem (NLP), 260

664

SUBJECT INDEX

Nonlinear variables, 355 Nonsingular matrix, 637 Null matrix, 635 Null vector, 652 Numerical gradient approximations, 129 Nurse scheduling, 517

OPTISEP penalty function program, 280– 281 OPTLIB, 281 OSL, 178, 204, 493 OTC, 179 Oxygen supply system design, 8

OBI, 178, 204 Objective function, 28 Opportunity cost, 183 OPRQ quadratic program, 470 OPT: application example, 627 comparison with other methods, 536 Optimal solution, 154, 156, 161 Optimal value, 154, 156, 161 Optimality criteria, 35, 80 Optimality criterion theorem, 187 Optimization applications, 6, 545 for analysis, 20 case study, 603, 610, 621 for data reduction, 20 in design, 8 for operations, 15 for planning, 15 Optimization methods: algorithm testing, 532 comparison criteria, 532, 538 comparison of, 531 Colville study, 533 Eason and Fenton study, 533 Sandgren study, 534 Schittkowski study, 537 surveys of, 533 efficiency criteria, 532 software testing, 532 testing of, 532 Optimization problems: constrained, 29 fractional linear program, 29 integer program, 29 linearly-constrained, 29 linearly-constrained nonlinear programs, 29 linear programs, 29 quadratic program, 29 single-variable, 29 unconstrained, 29 Optimization Technology Center (OTC), 531, 540 Optimization theory, 1 Optimum sloppiness, principle of, 544

Palacios-Gomez linear programming strategy, 462 Parabolic penalty function, 262, 265 Parallel subspace property, 101 extended, 104 Partitioning algorithm, 514, 581 Pattern search method, 310 Penalty function: algorithm for, 279 bracket operator, 264, 269 computer programs for, 280 equation, 261 exterior, 261 inequality penalty, 269 infinite barrier, 262 interior, 261 inverse penalty, 264, 273 log penalty, 263, 270 parabolic penalty, 262, 265 penalty parameter, 277 surveys of, 280 use of, 260 Penalty term, 261 Phase I problem, 176, 432 Piecewise linear approximation, 359 grid points, 359 separable function, 363 single-variable functions, 359 Pivot element, 169 Pivot operation, 162, 169 Pivot row, 169 Point-estimation, 55 Polak and Ribiere conjugate gradient method, 122 Polymer plant optimization, 545 Polynomial approximation, 55 Pooling problem, 356, 581 P␾P II: application of, 355 linearization program, 352 Portfolio selection, 494, 517 Positive definite, 499, 501 Positive definite matrix, 640, 642, 645, 650 Positive semidefinite, 499, 501 Positive semidefinite matrix, 640, 642, 650

SUBJECT INDEX

Post-optimality analysis, 180 Powell’s conjugate direction method, 106 Powell’s method, 59, 69 Powell’s search, 69 Preemptive goal programming, 509 Principal determinant, 641 Principal minor, 641 Principle of optimum sloppiness, 544 Problem formulation, 2 case study, 603, 610, 621 independent variables, 4 level of detail, 5 performance criterion, 3 system boundaries, 2 system model, 5 system parameters, 4 Problem implementation: model assembly, 553 procedure for, 552 Production planning, 517 Pseudo convex function, 250 Pseudo concave function, 250 Pure integer program, 481 Quadratic approximating function, 583 Quadratic approximation, 451. See also Constrained, quadratic approximation methods Quadratic convergence, 117 rate of, 107 Quadratic estimation methods, 56 Quadratic forms, 640 completing the square, 642 indefinite, 641–642, 645 negative definite, 640, 642, 645, 650 positive definite, 640, 642, 645, 650 positive semidefinite, 499, 640, 642, 650 Quadratic programming, 451, 481, 494 applications of, 494 complementary pivot method, 502 Kuhn-Tucker conditions, 498, 501 Quadratic search, 58 Quality control, 517 Quality engineering, 27–28 Quasi-Convex function, 251 Quasi-Concave function, 252 Quasi-Newton methods, 109, 123 Random search methods, 322 adaptive step-size search, 326 combinatorial heuristic, 327

665

direct-sampling procedures, 322 interval reduction sampling, 323 simultaneous, 322 Raoult’s law, 550 Ray solution, of complementary problem, 505, 507 Redlich-Kwong equation, 22 Reduced gradient: first-order necessary conditions, 409 Lagrange multipliers, 409 Lagrangian optimality criterion, 409 Reduced gradient formula, 409 Reduced gradient method, 399 algorithm for, 400 convergence acceleration, 403 convergence proof of, 402 jamming, 402 Reduced second derivative, 473 Redundant constraint, 565 Refinery planning, 16 Region elimination, 45 bounding phase, 46 elimination property, 46 fractional reduction, 54 interval refinement phase, 46, 48 Region elimination search, 69 Regression coefficients, 497 Relative minimum, 36 Relative profit coefficients, 165, 169 Requirement vector, 159 Response-surface model, 582 Restricted basis entry, 365 relaxation of, 367 Risk-averse model, 497 Robustness, of algorithms, 532 Rosen and Suzuki nonlinear program generation, 537 Rosenbrock’s function, 133 Row echelon form, 162, 644, 652 Row operations, 162 Row vector, 635

Satty’s priority theory, 538 Saddlepoint, 38, 80, 222, 235 Saddlepoint conditions, 235 existence of saddlepoint, 237 necessary optimality theorem, 237 sufficient optimality theorem, 236 Saddlepoint problem, 236 SALMNF: method of multipliers program, 293

666

SUBJECT INDEX

SALQDF: method of multipliers program, 293 SALQDR: method of multipliers program, 293 Sandgren study, 534 Scalar function, 32 Scaling: application example, 618 in model preparation, 555 Schittkowski nonlinear programming study, 537 Search methods, 45 Secant method, 64 Second-order necessity theorem, 241 Second-order optimality conditions, 238 Second-order sufficiency theorem, 243 Sectioning, 399 Sectioning technique for multivariable optimization, 92 SEEK1: penalty function program, 281 SEEK3: penalty function program, 281 Sensitivity analysis, 180, 590 opportunity cost, 183 shadow prices, 182, 225 Separable functions, 362 Separable programming, 359 iterative refinement, 368 linear programming solutions, 364 restricted basis entry, 365 Set theory, 652 Shadow prices, 182, 225 Simple gradient method, 110 Simplex method, 163 alternate optima, 171 artificial variables, 174 basic feasible solution, 162 basis change, 174 computational efficiency of, 179 computational steps of, 171 computer solution of, 177 condition of optimality, 166 cycling, 174 degeneracy, 174 inner product rule, 166 maximization problem, 171 minimization problem, 161 minimum ratio rule, 167 Phase 1, 176 Phase 2, 176 principles of, 161 relative profit, 165

two-phase, 176 unbounded optimum, 173 Simplex search method: calculations for, 88 modified, 90 for unconstrained multivariable optimization, 86 SIMPLX: penalty function program, 281 Simulation libraries, 554 Slack variable, 159 SLAM simulation model, 583 Small step gradient methods, 358 Software availability, 530 Software testing, 532 Solution evaluation: interpretation, 589 Lagrange multiplier values, 591 parameter case study, 591 post-optimality analysis, 168 use of simplified auxiliary models, 589 validation, 589 see also Sensitivity analysis Solver, 531, 540 Sparse matrix, 639 Square matrix, 635 Starting point generation: penalty-type function, 309 random, 309 sequential constrained minimization, 309 see also Feasible, starting point generation Static question, in optimization, 35, 78 Stationary conditions, 81 Stationary point, 38, 41, 81 Strict local minimum, 238 Strict optimum, 80 Successive linear programming: Kuhn-Tucker conditions, 355 step-size adjustment, 350 use of penalty functions, 352 Successive quadratic programming method, comparison with other methods, 539 Sufficient conditions, for multivariable function, 83 SUMT: penalty function program, 280 SUMT-IV: penalty function program, 592 Superlinear convergence, 69 rate of, 107 Surplus variable, 159 Symmetric dual, 185 System, 2

SUBJECT INDEX

Taguchi system, 27–28 Taylor series, 36, 39, 239 Tetraethyllead requirements, computation of, 369 Through circulation drying system, 479, 597 Transformation methods: characteristics of, 261 variable bound, 289 Transportation, 517 Trust regions, 350, 355 Two-phase simplex method, 176

Unbounded optimum, 157, 161, 173 Unconstrained multivariable optimization: Broyden-Fletcher-Goldfarb-Shanno method, 126 Cauchy’s method, 110, 113 comparison of methods, 130 conjugate gradient methods, 117 Fletcher-Reeves method, 120 memory gradient method, 121 Polak and Ribiere method, 122 Davidon-Fletcher-Powell method, 124 direct methods, 85 dynamic question, 79 gradient-based algorithm, 128 gradient-based methods, 108 Hooke-Jeeves pattern search, 93 Marquardt’s compromise, 116 modified Newton’s method, 115 modified simplex procedure, 90 necessary and sufficient conditions, 82 Newton’s method, 111 numerical gradient approximations, 129 Powell’s conjugate direction method, 106 quasi-Newton methods, 123 simplex search method, 86 static question, 78

667

Unconstrained multivariable problem, 78 Unconstrained optimization, of several variables, see Unconstrained, multivariable optimization Unimodal function, 34, 45 Unimodality, 35 Unique local minimum, 238 Unique optimal solution, 156 Unique optimum, 161 Univariate search: Bisection method, 63, 69 Cubic search, 65, 69 Exhaustive search, 54 Golden section, 51, 54–55, 69 Interval halving, 48–50, 54 Newton-Raphson, 61 Powell’s method, 59, 69 Quadratic search, 56 Secant method, 64 Unrestricted variables, 159 Variable elimination, 219 Variable metric methods, 123 for constrained optimization, 464 Vectors, 652 VF01A: method of multipliers program, 293 Water resources, 517 Watter’s transformation, 483 Weak duality theorem, 186, 201 Welded beam design, 11 Welded beam problem, 591 Wood’s function, 136, 291 XMP, 178, 204 Xpress-SLP, 355

Engineering Optimization, 2nd Ed, Wiley (2006)

Recommend Documents